Jeff Bezos throws his hat in the ring for an orbital data center megaconstellation, too

"Space-based data centers will be a complement to terrestrial infrastructure."

See full article...

 

[dmsilev]

It is not that difficult in reality. Especially if you add some heat pumps to increase the radiator temperature. Scott Manley has a nice video about it.

The problem is that it doesn't scale well. Manley's video is about a 20 kW load, which by the standards of data center compute is a teeny tiny little thing, well under one rack's worth of stuff. So, your choices are either to scale up each satellite by a factor of ten or a hundred, or to partition your distributed data center into lots and lots and lots of tiny nodes. Neither option is attractive. The latter approach will kill you on latency for any compute task that doesn't fit inside half a dozen GPUs. Which for modern AI is to first order "all of them". The former means that your individual nodes are a lot bigger, and also less efficient, The average distance from the load to the radiators goes up, and even with heat pipes and the like the efficiency of heat transfer goes way down. You almost have to adopt the ISS approach of active coolant loops flowing into the radiators, which means fun with pumps and so forth in a completely zero-maintenance context. Good luck!

 

[/or\]

Just when you thought the data center bubble was going to burst it gets bigger.
How many teraflops is in a flop?

 

[BC_Sizemo]

Are there any international laws in space? Obviously, you have left the regional / country level laws behind, much like going into international waters. But are there global laws that are just banned out right? If a country decides to limit the development of AI I can easily see how that basically throws a monkey wrench in any data centers located there. Which in turn limits your development speed of any AI system. Where as in space....you can just let AI run free and see what happens.

*and when I say AI I mean something beyond an LLM.

 

[frogstomp]

Oh fuck off, Jeff.

 

[Malmesbury]

Can we stop pretending that the FCC chairman, who has for years been advocating for SpaceX and accused the previous FCC of "regulatory harassment" against them, is some principled independent arbiter here?

The following can be true, at the same time

1) The FCC chairman is a classic Trump appointment. With all that implies.
2) The Sue Origin moniker was earned long before this. And deserved.

 

[caramelpolice]

It's cool how much of the AI bubble is caused by complete morons just spouting blatant nonsense like 'our AI models are gonna take everyone's jobs!' or 'we're gonna put data centers in space!' while the press just regurgitates it uncritically and investors' eyes light up like a baby having keys jangled over their crib.

 

[Tactical Finesse]

Are there any international laws in space? Obviously, you have left the regional / country level laws behind, much like going into international waters. But are there global laws that are just banned out right? If a country decides to limit the development of AI I can easily see how that basically throws a monkey wrench in any data centers located there. Which in turn limits your development speed of any AI system. Where as in space....you can just let AI run free and see what happens.

*and when I say AI I mean something beyond an LLM.

Oh, yes. It dates back to the 60s. The problem with such treaties--it is up the good-will and honor of good people to respect them. And when they don't--it is just words on paper.

"Key provisions of the treaty include prohibiting nuclear weapons in space; limiting the use of the Moon and all other celestial bodies to peaceful purposes; establishing that space shall be freely explored and used by all nations; and precluding any country from claiming sovereignty over outer space or any celestial body. Although it forbids establishing military bases, testing weapons and conducting military maneuvers on celestial bodies, the treaty does not expressly ban all military activities in space, nor the establishment of military space forces or the placement of conventional weapons in space.[7][8]

The OST also declares that space is an area for free use and exploration by all and "shall be the province of all mankind". Drawing heavily from the Antarctic Treaty of 1961, the Outer Space Treaty likewise focuses on regulating certain activities and preventing unrestricted competition that were thought might lead to conflict at that time.[9] Consequently, it is largely silent or ambiguous on newly developed space activities such as lunar and asteroid mining.[10][11][12]"

https://en.wikipedia.org/wiki/Outer_Space_Treaty

 

[Nooge]

It is not that difficult in reality. Especially if you add some heat pumps to increase the radiator temperature. Scott Manley has a nice video about it.

I design air conditioning and heat pump systems for a living. They are great because at many operating conditions you can reject 2-3 kW of heat for every 1 kW of energy you input.

BUT, that’s only because we ignore the heat that earth’s enormous atmosphere absorbed. If you actually account for all the energy, you’ll see that you had to add heat to the system to compress the refrigerant. So now you have more heat to reject than before. So using a heat pump is just making the problem of poor heat rejection worse.

It’s better to just use a coolant loop to move the heat to the radiators with minimal power usage and added heat. Heat pumps and air conditioning are only needed when you need to get colder or hotter temperatures than are available in the ambient environment.

Edit: I just saw your other reply about raising the temperature to increase the radiative (T^4) losses. I suppose with careful selection of the refrigerant and optimization for a single operating condition (space is always the same temperature), that could make sense. Computer chips tend to not like running above 50C or so. Let’s say the right refrigerant gets up to 200C at the condenser, that’s (473K^4 / 323K^4) = 1.464 times better heat transfer, which is worthwhile but not drastic, especially since you now have probably 20-30% more heat to transfer. If they’re able to close to 300C at the condenser, then it’s a big win.

 

[Dan Corcoran]

If you want a deeper engineering perspective on this, Scott Manley just put out a podcast (“Is It Really Impossible To Cool A Datacenter In Space?” — March 19, 2026) that’s directly relevant.

The key takeaway is why small is better: ~20 kW is fairly doable, ~100 kW is pushing it, and beyond that the radiator scaling problem and maintaining orientation becomes brutal.

It really highlights the reasons behind the current thinking to have tens of thousands up to a million small orbital data centers.

 

[Doomlord_uk]

What does anything that happens in space have to do with the FCC anyway? Just curious.

 

[Tactical Finesse]

They don't have to put entire racks in a single satellite. The laser interlinks (not radio) mean that bandwidth between these units will be plentiful and low latency.

The "depreciation schedule" on this hardware is based on terrestrial data center opportunity cost and ongoing operating expenses. On Earth, you don't want to run 5-year-old hardware because the electricity, cooling, and datacenter space could be better spent on newer and more efficient hardware. Yet as I've said in other comments in other threads, this math is wholly different in space.

Once you have launched the satellite, you have paid for it's lifetime power and cooling needs. The only ongoing expenses are whatever the network downlink costs (trivial when you own the network) and any constellation management costs related to collision avoidance. Keeping older hardware running until their maneuvering fuel is low in order to run slower models (or even newer models more slowly) means you can keep eeking out revenue.

Think of how semiconductor manufacturing companies will keep running older fabs to produce less advanced chips, because the critical costs are the up front hardware and not the ongoing operations. The cost of putting datacenters in space is likely to be pretty comparable to building a new datacenter on Earth once you actually factor in the zoning politics, build timelines (the satellites can be "mass-produced" in a factory rather than built and wired on site), and lifetime power and cooling costs.

Space datacenters are going to be the only way to scale compute in a political environment that is increasingly hostile to new terrestrial datacenters. Ars doesn't give it a lot of attention, but there are a ton of jurisdictions throughout the US which are actively slowing or blocking construction of datacenters and considering legislation to outright ban them. There's even talk from members of Congress about moratoriums. They're noisy, they use lots of electricity (either driving grid prices up for constituents or burning fossil fuels and polluting heavily to generate their own supply), and often use a lot of water for cooling.

You don't have to put entire racks in a single satellite--but then you create other problems. Like needing 100,000+ satellites to do the same compute as a single terrestrial data center (depending on the size of "data center' you want to compare to since they're all different).

 

[Varste]

I completely agree that the whole thing seems utterly pointless, but maybe I am just too stupid to understand thinking on their level.

Unless your raison d'etre is to make more money than any human could ever spend in their life time, then no you're not on their level and you should be all the happier for it. These people sole existence is to make line go up. How fulfilling.

 

[PERSJWM]

Man, to hell with these people.

 

[EllPeaTea]

What does anything that happens in space have to do with the FCC anyway? Just curious.

Because the things in space need to communicate with things on the ground. And that requires a radio licence.

 

[SunnyD]

Putting all this shit in space all I can think about is this scene.


View: https://www.youtube.com/watch?v=92LDtH2ZB-w

 

[Statistical]

It was dumb when SpaceX proposed it. It is still dumb when BO does the same thing.

All of this comes with serious handwavium that somehow it will make sense and work and be profitable. It is just like Amazon and Google and the rest of big tech dumping over a trillion dollars into terrestrial datacenters that will somehow make a profit someday. At least terrestrial datacenters can be used for other things and servers replaced easily. There will likely be a glut but non-AI demand will eventually soak it up. Datacenters in space don't even have that lemonade from lemons fallback option.

The "plans" all remind me of this

 

[**RTS**]

It is not that difficult in reality. Especially if you add some heat pumps to increase the radiator temperature. Scott Manley has a nice video about it.

Bigger issue is all the trash being lifted up there and then deorbited. But when have rich ever bothered with impact of their BS on others. If they cared about others they would not be rich.

Watched the video. But the big issue is that he is way underestimating the power consumption of an AI datacenter. 20-100kW is nothing when AI datacenters are pushing 144kW per rack, of which a datacenter has many.

 

[floyd42]

Don't have to obey any laws in space.

 

[Statistical]

Ars is the perfect publication to either explain how these space-based AI data centers make any financial sense, or to show that it’s absolutely stupid (my bet is on the latter). Frankly, I’m disappointed that there hasn’t already been an investigative article on this, instead of just parroting the space company press releases.

Yeah it is all handwavium. Can you put a datacenter in space? Sure. You can put a Lexus Dealership in space too. It doesn't mean it makes any sense.

Just like the terrestrial AI datacenter buildout nobody is making a single cent (except component suppliers like NVidia) but all of big tech keep just throwing endless mountains of cash into the fire in order to not be left behind. Just on Earth we have now passed the $1T mark in terms of datacenter for AI mark that somehow based on handwavium will someday turn a profit. The plans over the next 5 years are another $2T on top of the current $1T.

 

[famousringo]

They need to make a second Doomsday Clock, but this time for Kessler Syndrome rather than nuclear war.

They just made it. It's called the CRASH clock:

https://outerspaceinstitute.ca/crashclock/

 

[artimusprime]

If you want a deeper engineering perspective on this, Scott Manley just put out a podcast (“Is It Really Impossible To Cool A Datacenter In Space?” — March 19, 2026) that’s directly relevant.

The key takeaway is why small is better: ~20 kW is fairly doable, ~100 kW is pushing it, and beyond that the radiator scaling problem and maintaining orientation becomes brutal.

It really highlights the reasons behind the current thinking to have tens of thousands up to a million small orbital data centers.

I'm curious so I'll check that pod out but does it get into things like high availability, hardware management. like maybe if you took away the need to replicate data from one center to another it'd get easier, but it really just seems like there's a lot of challenges for not much benefit.

 

[TauCeti]

Guess they also will sell shade (from the constellation) to the highest bidder on earth...

 

[jmellott]

What's the play here? I don't understand what the benefit of a data center in the sky is over a data center in Stuttgart, Arkansas.

Or is this just a ploy for investor dollars and to justify Bezos' space ambitions?

 

[fenris_uy]

Watched the video. But the big issue is that he is way underestimating the power consumption of an AI datacenter. 20-100kW is nothing when AI datacenters are pushing 144kW per rack, of which a datacenter has many.

And the constellation has many satellites.

Asume that BO can only get 80kW of compute per sat (20kW reserved for communication, station keeping and loses), that is still 4 GW of compute in the constellation. It's not ground breaking, but it's not that little also.

 

[Statistical]

Are there any international laws in space? Obviously, you have left the regional / country level laws behind, much like going into international waters. But are there global laws that are just banned out right? If a country decides to limit the development of AI I can easily see how that basically throws a monkey wrench in any data centers located there. Which in turn limits your development speed of any AI system. Where as in space....you can just let AI run free and see what happens.

*and when I say AI I mean something beyond an LLM.

Well first there really is no such thing as international law at least in the sense that people think of domestic laws. Countries are sovereign. Counties may voluntarily enter into treaties with other countries and this forms the basis of international law but it isn't like there is an international police force and international. Countries can also exit treaties or simply ignore them. Other countries can take actions like sanctions to discourage that but there is no "the US broke the international law on X so the international army will force the US to stop".

The Outer Space Treaty (OST) of which the US is a signer says the US has an obligation to regulate activities in space by parties using equipment launched to space by the US. To name a specific example as long as child porn generation is illegal in the US then it is illegal in US operated datacenters in space too.

There is no you put it in space and laws don't apply provision.

 

[Statistical]

What's the play here? I don't understand what the benefit of a data center in the sky is over a data center in Stuttgart, Arkansas.

Or is this just a ploy for investor dollars and to justify Bezos' space ambitions?

Companies don't want to get left behind on AI and as a result are pissing away fortunes so in case someone does somehow figure out how to make money they won't be behind. It is this self perpetuating cycle. Company X spend $100B so company Y goes we can't be left behind and spends $150B. There is no guarantee any of them will make any money in fact nobody even has a hard nuts and bolts plan on where the money will come from they just want to be ready in case it does.

At least with BO there is currently only one shareholder so Bezos is just lighting his own money on fire.

 

[EllPeaTea]

What's the play here? I don't understand what the benefit of a data center in the sky is over a data center in Stuttgart, Arkansas.

Or is this just a ploy for investor dollars and to justify Bezos' space ambitions?

It's going to get harder to get permits for big data centres on the ground. Isn't Ohio looking to ban them entirely? And then you need to supply them with power, which necessitates grid upgrades, or colocating gas turbines.

Data centre in the sky brings its own power supply (via solar panels), and no need to deal with local or state politics.

 

[Deranged]

Heat pumps are not magic. The limit isn't even moving the heat from compute hardware to the radiator. The limit is how quickly the radiator can radiate the heat into space. You're going to quickly saturate the radiator unless it's massive.

It doesn't answer the fundamental question of why you even need this in the first place. Communication satellites solved a problem that has plagued humanity since it first started organizing into civilizations. There is no alternative to satellite comms when you are away from terrestrial networks. It's why comm satellites were one of the first practical applications of space.

Data processing doesn't have that need. If you have a network, you don't particularly care about where the numbers are being crunched. The data will get back to you. Which is why data processing has, typically, been done on the ground and why spaceborne data centers haven't really been attempted. It's not that we can't, it's that there's no point.

I suppose a possible idea is that datacenters in space wouldn’t be subject to all those pesky privacy and data protection laws that exist on Earth? However, that seems like it wouldn’t make much difference as these datacenters need to connect to ground based infrastructure where those laws DO exist before reaching recipients anyway.

Other than that, I got nothing. A guy whose work to a large part consists of datacenters and rocket launch would be pretty intimately in tune with the cost and requirements of both, so possibly this is just someone blowing smoke… somewhere?

 

[terkans]

Guess they also will sell shade (from the constellation) to the highest bidder on earth...

Not from those "Terminator Sun-synchronous orbits", their shade doesn't fall on earth, for the same reason that the satellites are always in sunlight. Terminator here means the circle where sunlight turns to shade.

The methaphical shade, on the other hand, will fall on all of us.

 

[Jharm]

Ars is the perfect publication to either explain how these space-based AI data centers make any financial sense, or to show that it’s absolutely stupid (my bet is on the latter). Frankly, I’m disappointed that there hasn’t already been an investigative article on this, instead of just parroting the space company press releases.

This shouldn’t be hard. I feel like anyone in the space industry should be able to do some back-of-the-napkin math to show that it doesn’t make sense even with very generous assumptions (i.e., assume Starship is working and launches cost roughly what they’re currently projecting).

Some things to include:

• you have to compare against terrestrial data centers with the same latency (distance) to cities as low earth orbit
• whatever magic hand-waive they want to do for collecting heat (to make waterless cooling work in space) is also available for terrestrial data centers. Current data centers “need” water for open-loop cooling because that’s very cheap, but obviously they won’t have that in space.
• at least mention the problem of handing off user sessions, since the data center that was “close” to the user a few minutes ago is now over a different continent; for long AI queries, this can happen between when the user asks the question and when an answer is finished.
• AI data centers “need” lots of grid power, but they’ll be forced to provide their own power in space. Compare that price to them providing their own power for terrestrial data centers (yes, solar panels in space can get sunlight 24/7, but I suspect considerably cheaper to install panels on earth)

but that is the reason. Solar panels are dirt cheap but does not work 24/7 on earth. In space they do. So you have to compare what is the cost of electricity on earth vs transport cost to a place where electricity is free.

So if electricity is on average is 17 cents pr
kWh. And let us assume you use 20kW for 5 years. Then the electricity bill is approximately 150k usd. Not including any cooling power.

So if you build and transport that 20kW rack to space and last for 5 years for less than 150k then you could argue there might be a business case on electricity alone.

Edit:
If 60 satellites pr star ship launch and cost is 10 million or launch gives a cost or satellite to 160k. So napkin math shows same order of price.

 

[fellow human]

Excellent article from a bonafide expert on such things about how stupid this idea is:

https://taranis.ie/datacenters-in-space-are-a-terrible-horrible-no-good-idea/

 

[lolnova]

They don't have to put entire racks in a single satellite. The laser interlinks (not radio) mean that bandwidth between these units will be plentiful and low latency.

No it won't.

On Earth, you don't want to run 5-year-old hardware because the electricity, cooling, and datacenter space could be better spent on newer and more efficient hardware. Yet as I've said in other comments in other threads, this math is wholly different in space.

Correct, it's far more expensive in space.

The cost of putting datacenters in space is likely to be pretty comparable to building a new datacenter on Earth once you actually factor in the zoning politics, build timelines (the satellites can be "mass-produced" in a factory rather than built and wired on site), and lifetime power and cooling costs.

Lol.

 

[Statistical]

but that is the reason. Solar panels are dirt cheap but does not work 24/7 on earth. In space they do. So you have to compare what is the cost of electricity on earth vs transport cost to a place where electricity is free.

So if electricity is on average is 17 cents pr
kWh. And let us assume you use 20kW for 5 years. Then the electricity bill is approximately 150k usd. Not including any cooling power.

So if you build and transport that 20kW rack to space and last for 5 years for less than 150k then you could argue there might be a business case on electricity alone.

Solar panels may be cheap but putting a solar panel into space isn't cheap not even when cheapER launch vehicles. Solar panels in space produce roughly 4x to 7x the annual output compared to on the Earth depending on exactly where you place them. The cost of delivery is on the order of 10,000x higher.

As a thought exercise if the power aspect of datacenters in space make sense why not just build solar power farms in space and beam it datacenters (and other high demand consumers) on the ground? If space based power doesn't make sense because it can't compete with terrestrial power prices how does attaching a datacenter to your uneconomical power generation become economical?

 

[Jharm]

but that is the reason. Solar panels are dirt cheap but does not work 24/7 on earth. In space they do. So you have to compare what is the cost of electricity on earth vs transport cost to a place where electricity is free.

So if electricity is on average is 17 cents pr
kWh. And let us assume you use 20kW for 5 years. Then the electricity bill is approximately 150k usd. Not including any cooling power.

So if you build and transport that 20kW rack to space and last for 5 years for less than 150k then you could argue there might be a business case on electricity alone.

Edit:
If 60 satellites pr star ship launch and cost is 10 million per launch that gives a cost or satellite per 160k. So napkin math shows same order of price.

 

[huskcummerbund]

They just made it. It's called the CRASH clock:

https://outerspaceinstitute.ca/crashclock/

Oh.... Well then... Nevermind. Thank you

 

[DarthSlack]

What's the play here? I don't understand what the benefit of a data center in the sky is over a data center in Stuttgart, Arkansas.

Or is this just a ploy for investor dollars and to justify Bezos' space ambitions?

It's exactly the same kind of business acumen that leads to "We lose money on every sale but we'll make it up in volume".

 

[fenncruz]

Don't have to obey any laws in space.

But the meatbags left on Earth supporting the data centers in space still have to follow the law.

 

[Jharm]

Solar panels may be cheap but putting a solar panel into space isn't cheap not even when cheapER launch vehicles. Solar panels in space produce roughly 4x to 7x the annual output compared to on the Earth depending on exactly where you place them. The cost of delivery is on the order of 10,000x higher.

As a thought exercise if the power aspect of datacenters in space make sense why not just build solar power farms in space and beam it datacenters (and other high demand consumers) on the ground? If space based power doesn't make sense because it can't compete with terrestrial power prices how does attaching a datacenter to your uneconomical power generation become economical?

I tried to put some numbers in my post. As you indicate it is all about cost. So if you can make a cheap transporter there might be a business case.

 

[floyd42]

But the meatbags left on Earth supporting the data centers in space still have to follow the law.

Sure, just like they do today.

I assume that raiding a data center in space would be a lot harder and pulling the plug on any of them will be just as hard, aka, not something us proles can do.