Don’t build the cloud above the clouds*
*TLDR don’t build server farms in space willy nilly, it would cost way too much. This does not include military/emergency infrastructure or computing capacity that is augmenting various recon and communications satellites. (cutting myself off otherwise this will get too long)
In the year of our yaddah yaddah 2kTwenty-Six, the push for “AI” dominates discussions for planning the future of economies of nations. The energy demands for “AI” data centers is becoming so intense that a team at Google wrote a white paper exploring building server farms in orbit.
Building stuff in space is awesome,* from satellites that let you download videogames in the middle of the ocean, to satellite dishes with hundreds of channels, to measuring our environment from space, should we add large scale compute to the list?
Time to do some math.
We have a selection of variables that we need to choose some values for. I want to give the concept the best possible chance of getting funding from whichever source of funding I am trying to convince to help get our mission into space.
We have our launch costs, satellite costs amortized cost to design and build, as we are trying to be a profit driven initiative, how much revenue will these spacecraft produce each day? This is a best-case scenario; we have no depreciation of our assets or system failures that lead to our spacecraft being inoperable.
Assume the following
Satellite Hardware: $1000/kg (cheap compared to the $11,000 it costs to buy a cube satellite kit that masses 1.33 kg)
Launch costs: $100/kg
Satellite specifications. 1 m^2 solar collector system producing 420 Watts of power *24/7/365, with compute modules that can use close enough to 100% of the power for processing data, as the energy spent transmitting data is trivial enough to be ignored.
Wow this is awesome.
That means every day our $1100 spacecraft produces 10.08 kWh of electricity.
We then need to provide a conversion rate of how much valuable computing time can be provided for every kilowatt hour.
If we assume that our computing time provides $0.20/kWh, our satellite produces $2.12** each and every day.
This means our $1100/kg satellite material and launch cost could be paid back in only 546 days ***. Super impressive ….
…. Until you think about things for more than 5 min (and with how my brain hyper focuses I want to be clear I spent far too much time on this question (not that I didn’t likely make mistakes, I am human and limited to what is publicly available as I don’t want to spend money that I could use for magic cards or other points of hyperfixation)
First assumption. $100/kg launch costs are beyond generous, even in the Google White paper they are assuming $250/kg to launch their satellite. No biggy our magic satellite set up now cost $1250/kg to build and launch. Our payback period has grown to 620 days, which could be worse (oh it will get worse).
The Solar Array Wings on the International Space Station are an amazing feet of engineering, massing 1088 kg they produce 31 kilowatts of power, or about 28.5 watts of power per kilogram of mass of the system. Huh, that’s weird, we assumed a single kilogram of magic satellite could produce 420 watts of power, as well as computing, and all the other shiny stuff our satellite needs. OK, looks like we need to revise our mass estimates, we can be generous, we aren’t launching our satellites tomorrow, new photovoltaic technologies should become available. Let us assume that we can achieve 100 watts per kilogram of power production while still allowing for all of the fun features we believe our theoretical satellite can achieve. Now our satellite system can only produce 2.4 kWh per kilogram launched.
The mission is looking rough, our 20 cents of compute time value per kilowatt hour of electricity available, in tandem with our $250/kWh launch cost has trashed our payback period. Now we are looking at 2,604 days to make back our money, over seven years.
It gets even worse if we look closer at the current estimates of how much value of computer processing time can be derived from a kilowatt hour of electricity. From the publicly available sources I could find, 20 cents is generous. According to Pew research cloud computing and the AI boom consumed 4% of US electricity production.
Wikipedia put US electricity production in 2023 at 4,104 trillion kWh. Translating to datacenters and cloud computing using about 164 billion kWh of electricity. If we take Yahoo finance at their word cloud computing was worth let us say a rounded up a trillion dollars in 2025. Dividing cloud compute annual revenue by annual energy consumption, we can give ourselves a loose number for the compute value from 1 kWh of electricity. $1 Trillion/164 billion kWh or $0.0609 kWh. Woof one kilowatt hour of electricity is only worth 6 cents of computer time.
Woah, that means that even if we triple the supposed value of compute time, our initial estimate of compute time being worth 20 cents per kilowatt hour was too high. The payback period is approaching 8 years, before we account for spacecraft losses or depreciation of the value of what is being done on our satellites. A facility on earth can swap out bad components and upgrade in a timely manner, a satellite has a harder time being repaired, and designing our satellite to be repaired makes it liable to cost more to build (I presume a repairable satellite design will cost more as you need better access points and the like, but as I am not a subject matter expert I will not speak definately about it.
Obie I hear myself saying, I thought you were pro space, this article seems like quite the downer on orbital cloud computing.
Well it is, but only with respect to moving computers off the surface of the Earth, while still asking those computers to do work on Earthy projects.
I do believe that orbital server farms that serve other satellites, that act as data relays, and/or pre-processors for data being sent down to Earth, those options can be very valuable and the math works out differently. A computer system that ensures that a forward operating base or rescue command center on Earth, the value of the processing capacity and responsiveness is very different than commodity computations being done in a static server farm. Sustainably building computers and factories in space to make space a more valuable resource for the people of Earth is valuable.
I hope this was interesting, I did try to keep my personal biases in check, and I do believe that I was as fair as possible to the "business case” for orbital commodity cloud computing. If folks have papers further reading they would recommend I welcome the feedback.
*citation needed
** rounded from 10.08kWh*$0.20 /kWh= $2.1168