My recent excursion into putting data centers in space brought an interesting fellow into my orbit.
His name is Shubber Ali.
He has been involved with the space industry since the 1990s, when he managed KPMG’s space consulting practice. For nearly 20 years, off and on, he has also written a blog called The Space Cynic.
This is not a man burdened by excessive enthusiasm.
Last November he wrote a piece with a title that leaves relatively little room for misunderstanding:
Why Space Data Centers Are Bullshit.
Subtlety has its place. Apparently not in headlines.
I had written my own, considerably more polite examination of the subject after hearing somebody suggest that the answer to the enormous power, water and land demands of terrestrial data centers was obvious:
Why don’t we just put them in space?
So I tried to build one.
Not literally.
I don’t have that kind of garage.
I took a representative one-megawatt orbital data center and followed the engineering.
That exercise gave us an estimated 44 tons of machinery, roughly 60,700 square feet of solar panels, about 26,900 square feet of radiators, heavy-lift rockets, liquid cooling, laser communications, robotic servicing, replacement parts and eventually the problem of getting the whole contraption safely out of orbit.
All of that for one megawatt of computing capacity — something that on Earth is relatively modest compared with today’s hyperscale data centers.
Shubber read my piece.
I read his.
We had arrived at much the same destination by different roads.
Then he sent me a note.
And introduced me to another problem.
Arithmetic.
The million-satellite solution
The orbital data-center idea has evolved.
The early visions of giant platforms containing enormous amounts of computing power invite obvious questions about mass, heat, construction and servicing.
SpaceX is publicly describing another architecture.
SpaceX itself acknowledges the scale of the problem. In its May 2026 securities filing with the Securities and Exchange Commission, the company said deploying 100 gigawatts of orbital computing capacity each year would require thousands of Starship launches annually and moving approximately one million metric tons into orbit every year.
That’s not a critic’s estimate.
That’s SpaceX’s.
The company also describes a long-term goal of launching 100 gigawatts of computing capacity into space each year.
One hundred gigawatts.
That’s 100,000 megawatts.
So we’re no longer talking about my little one-megawatt machine.
Instead of building one enormous data center, distribute the computing among enormous numbers of satellites.
A swarm.
A very large swarm.
Perhaps a million satellites.
This is where Shubber reached for a calculator.
Calculators have ruined many perfectly good ideas.
Let’s be generous
What I like about Shubber’s exercise is that he doesn’t begin by assuming the proposal fails.
Quite the opposite.
He gives it some remarkably favorable assumptions.
First:
Assume Starship works.
Starship has not yet demonstrated the routine commercial orbital transportation system required for an undertaking like this.
But we’re trying to find out whether the idea works, not win an argument.
So give them Starship.
Second:
Assume a Starship can carry about 90 data-center satellites per launch.
Shubber gets there by assuming roughly three times the satellite capacity of a Falcon 9 launch, while allowing that computing satellites would probably be larger than today’s Starlink satellites.
That’s not a SpaceX specification.
It’s Shubber’s assumption for doing the arithmetic.
Fine.
We’ll give them 90.
Then Shubber gets downright charitable.
He assumes SpaceX can launch Starship:
500 times a year.
Five hundred.
That works out to roughly:
42 launches a month.
Almost:
10 launches every week.
About:
1.4 launches every day of the year.
Christmas.
Thanksgiving.
Hurricanes.
Lightning.
Maintenance.
Technical problems.
The occasional rocket deciding it has other plans.
Five hundred.
For perspective, SpaceX’s record Falcon launch cadence in 2025 was roughly 165-plus flights.
That itself was an extraordinary achievement.
Shubber has effectively handed the hypothetical Starship operation about three times that annual launch rate.
This is less like stacking the deck against SpaceX than dealing it a royal flush.
Now let’s play the hand.
90 times 500
Ninety satellites per launch.
Five hundred Starship launches per year.
That gives us:
45,000 data-center satellites placed into orbit every year.
Forty-five thousand satellites a year would be an astonishing industrial achievement.
So how long would it take to launch one million?
The simple arithmetic says:
1,000,000 ÷ 45,000 = 22.2 years.
Twenty-two years.
Start in 2030 and, sometime around 2052, the millionth satellite goes up.
Champagne all around.
Except that’s not actually what happens.
That calculation quietly assumes something rather important:
Every satellite we launch is still operating when we finally launch satellite number one million.
Under Shubber’s assumptions, it isn’t.
And that’s where this gets interesting.
The first satellites are getting old
Shubber assumes these low-Earth-orbit computing satellites have useful lives of roughly five to seven years.
That number needs some qualification.
A satellite doesn’t encounter its fifth birthday and immediately fall out of the sky from embarrassment.
Spacecraft can be designed to operate considerably longer.
But five to seven years is a reasonable assumption for this exercise, particularly when we’re talking about rapidly evolving computing hardware in low-Earth orbit and processors that may not have the radiation hardening of traditional long-life spacecraft.
There is also another kind of death that matters here.
Obsolescence.
The spacecraft might work perfectly.
The computer aboard it may simply no longer be worth operating.
Anybody with an old computer or iPhone sitting in a drawer understands the principle.
It works.
You just don’t particularly want to use it.
AI processors are advancing rapidly. A GPU doesn’t have to physically fail to become economically obsolete.
So let’s use Shubber’s five-year assumption and see what happens.
We never get to a million
Year One:
45,000 satellites.
Year Two:
90,000.
Year Three:
135,000.
Year Four:
180,000.
Year Five:
225,000.
We’re not even one-quarter of the way toward our million-satellite data center.
Then Year Six arrives.
We have another 500 Starship launches available.
Those rockets can carry another 45,000 satellites.
Excellent.
Except the 45,000 satellites we launched in Year One have now reached the end of their assumed five-year useful lives.
So those 500 Starship launches are needed for two jobs.
We need them to launch the next 45,000 satellites required to continue building the constellation.
And we need them to launch 45,000 replacements for the satellites from Year One.
We can’t do both.
Not with 500 launches.
If we replace all 45,000 aging satellites, we’ve consumed our entire annual launch capacity merely keeping the existing constellation operating.
We add:
45,000 new satellites.
And lose:
45,000 old satellites.
Net gain:
Zero.
Year Seven arrives.
Now the satellites launched in Year Two need replacing.
Another 45,000.
Another 500 launches.
Year Eight?
Year Three’s satellites.
And so it continues.
Under these assumptions, we don’t reach one million satellites in 22 years.
We never reach one million satellites at all.
The constellation effectively tops out around:
225,000 active satellites.
The reason is wonderfully uncomplicated:
45,000 satellites launched per year × five years of useful life = 225,000 satellites.
That’s approximately all the launch system can sustain.
After the fifth year, we’re running just to stay where we are.
The 22-year figure turns out not to be the construction schedule.
It’s the construction schedule for satellites that don’t age.
Unfortunately, we’ve yet to develop those.
And that’s before anything breaks early
Our calculation has been remarkably polite.
It assumes every satellite lasts its full five years.
Space is unlikely to be equally courteous.
Some satellites will fail early.
A power system quits.
A communications system dies.
A processor fails.
A satellite loses propulsion.
A solar array doesn’t deploy properly.
A cooling system develops a problem.
Something gets damaged.
Those failures don’t check the construction schedule before occurring.
So now, while rockets are still trying to build the original million-satellite constellation, other rockets have to carry replacements for satellites we’ve already launched.
Perhaps some missions also carry replacement modules or servicing equipment.
We now have two launch programs competing for the same rockets:
Construction launches to build the data center.
Replacement and repair launches to keep the part we’ve already built operating.
Every rocket carrying replacements is a rocket that isn’t carrying satellites needed to finish the original constellation.
Unless we increase the number of launches.
But remember where we started.
Shubber had already given SpaceX the extraordinarily generous assumption of:
500 Starship launches every year.
Nearly ten every week.
Year after year.
And under the five-year satellite-life assumption, even that isn’t enough to finish the million-satellite constellation.
We are replacing the first computers while we’re still trying to launch the last ones.
At some point this stops looking like a construction schedule and starts looking like painting the Golden Gate Bridge.
Except the paint cans arrive by rocket.
So how many launches would it take?
Let’s turn the question around.
Suppose we somehow get all one million satellites into orbit.
Now we have to keep them there.
At a five-year useful life, approximately one-fifth of the constellation needs replacing every year.
That’s:
200,000 replacement satellites annually.
Not to expand the data center.
Not to increase its computing capacity.
Not to build another constellation.
Just to keep the existing one from shrinking.
At 90 satellites per Starship:
200,000 ÷ 90 = about 2,222 Starship launches every year.
That’s roughly:
185 launches every month.
About:
43 launches every week.
More than:
Six Starship launches every day.
Monday.
Six rockets.
Tuesday.
Six more.
Wednesday.
Keep going.
Christmas.
Thanksgiving.
Hurricanes.
Lightning.
Maintenance.
Technical problems.
Launch failures.
Six a day.
Just replacing computers.
And remember the assumption that seemed almost absurdly generous when we started?
500 Starship launches per year.
A mature million-satellite constellation with five-year hardware now needs more than 2,200 launches per year simply for replacement.
More than four times our original launch assumption.
And we haven’t added growth.
We haven’t added emergency replacements.
We haven’t added repair missions.
We haven’t added launches for anything else SpaceX might want to do.
Moon missions.
Mars missions.
Commercial satellites.
Military payloads.
NASA missions.
Other customers.
Our rockets have acquired a full-time job.
They’re changing the computers.
Fine. Make them last ten years
Maybe five years is unfair.
Let’s double it.
Give every computing satellite a useful life of:
Ten years.
Now only one-tenth of the million-satellite constellation needs replacing each year.
That’s:
100,000 satellites annually.
At 90 satellites per Starship:
About 1,111 Starship launches every year.
Roughly three launches every day.
Still more than twice Shubber’s original assumption of 500 launches per year.
And notice what happens during construction.
At 45,000 satellites launched annually, after ten years we’ve deployed:
450,000 satellites.
Then the satellites from Year One begin reaching their ten-year replacement point.
Once again, the rockets needed to continue building the constellation begin competing with the rockets needed to replace the equipment already there.
At 500 launches a year, a ten-year satellite life therefore doesn’t get us to one million either.
The rough sustainable ceiling becomes:
45,000 × 10 = 450,000 active satellites.
Better.
Still less than half the million.
All right. Fifteen years
Give the satellites 15 years.
At 45,000 satellites per year, the rough sustainable population becomes:
675,000 satellites.
Still short.
To reach one million satellites at a deployment rate of 45,000 per year without increasing the launch rate, the average satellite needs to remain useful for at least:
1,000,000 ÷ 45,000 = about 22.2 years.
There’s our 22-year number again.
But now it means something different.
The satellite doesn’t merely need to be launched within the 22-year construction period.
The earliest satellites need to remain useful for roughly the entire construction period.
Twenty-two years.
For spacecraft carrying cutting-edge AI computing hardware.
Think about the computer you were using 22 years ago.
That takes us back to around 2004.
Windows XP was current.
The first iPhone didn’t exist.
Facebook had just been founded.
YouTube didn’t exist.
Nvidia hadn’t introduced CUDA, the software platform that helped turn its graphics processors into the engines of modern AI computing.
Now imagine somebody telling you that the computing hardware installed in 2004 would still be an economically useful part of a cutting-edge AI data center today.
That’s what a 22-year useful life begins to ask of the first satellites in this constellation.
The spacecraft might survive.
The solar panels might survive.
The communications equipment might survive.
But the computers?
That’s a very different question.
Spacecraft engineers want hardware that lasts.
The computer industry makes its living ensuring that it doesn’t.
And remember what all this gets us
There is one final number hiding behind all this arithmetic.
Shubber assumes each data-center satellite provides approximately:
5 kilowatts of computing power.
One million satellites therefore provide:
5,000,000 kilowatts.
Which is:
5,000 megawatts.
Or:
5 gigawatts.
That’s serious computing capacity.
But remember what all this engineering and launch infrastructure gets us.
One five-gigawatt data-center system.
One.
The million satellites aren’t replacing America’s data centers.
They aren’t replacing the world’s data centers.
Under Shubber’s assumptions, the entire million-satellite constellation amounts to one enormous distributed five-gigawatt computing system.
Want another five gigawatts?
Start building the next million satellites.
And here’s where SpaceX’s own ambitions become particularly interesting.
In its May 2026 securities filing with the Securities and Exchange Commission, SpaceX described a much larger goal: eventually deploying 100 gigawatts of orbital computing capacity per year.
SpaceX said reaching that scale would require satellites delivering more than 100 kilowatts of compute per metric ton, approximately one million metric tons launched into orbit annually and thousands of launches every year.
That’s not Shubber’s arithmetic.
That’s SpaceX describing the scale of its own ambition.
Shubber’s thought experiment gives us:
5 gigawatts.
SpaceX is talking about eventually deploying:
100 gigawatts every year.
Twenty times Shubber’s five-gigawatt example.
Every year.
At that point we aren’t really talking about a satellite program anymore.
We’re talking about creating an orbital industrial economy whose raw material happens to include an extraordinary number of rockets.
Sunlight may be free.
Space has apparently decided to charge handling.
And after all that arithmetic, there is one other small matter we haven’t dealt with yet.
We still have to cool the computers.
And we haven’t cooled anything yet
None of Shubber’s arithmetic makes the engineering problems from my first column go away.
Every one of those million satellites produces heat.
Space is a vacuum.
And that’s the problem.
The computers still produce heat. That heat has to be collected, moved away from the processors and ultimately rejected as infrared radiation.
So every computing satellite needs thermal management.
Heat pipes.
Radiators.
Perhaps pumped liquid cooling.
Some means of carrying heat from the processors to radiator surfaces large enough to send that energy away into space.
Now multiply that problem by a million.
Every satellite needs electrical power.
Solar arrays.
Power electronics.
Perhaps batteries.
Every satellite needs communications.
Navigation.
Attitude control.
Propulsion or another means of maintaining orbit.
Radiation protection or radiation-tolerant electronics.
And every pound of that hardware is a pound the rocket isn’t using to carry another GPU.
The data-center satellite isn’t a computer with solar panels glued to the top.
It’s a spacecraft.
A million times over.
Now make a million computers talk
This brings us to Shubber’s next question:
How do we get the data back down?
A data center exists to process and move information.
Terrestrial data centers are connected by enormous fiber-optic networks.
AI processors exchange staggering quantities of information internally.
In my earlier column, I described the communications problem this way:
Imagine building a 12-lane interstate highway inside the computer and connecting it to Earth with a two-lane bridge.
Putting a million computers in orbit doesn’t make the bridge wider.
It may simply provide us with a million cars trying to get on it.
Optical communications — lasers — could provide enormous bandwidth between satellites.
That technology is promising.
Spacecraft could communicate directly with one another without sending every bit down to Earth and back.
But if the purpose of our orbital AI infrastructure is to serve people on Earth, information still has to cross between orbit and the ground.
Shubber put it more simply:
What’s the point of AI in space if it isn’t talking with Earth-based users?
There is a certain ruthless efficiency to a good question.
A million satellites need a network
So now build the communications architecture.
A million computing satellites.
Talking among themselves.
Routing information through optical links.
Connecting to ground stations.
Optical ground stations have a minor vulnerability.
Clouds.
Apparently nobody has yet persuaded clouds to respect venture-capital projections.
So the ground stations need geographic diversity.
Arizona.
California.
Hawaii.
Australia.
Europe.
Wherever the weather, geography and orbital paths make sense.
Radio communications have their own constraints.
Spectrum is finite.
Antennas have capacity.
Ground stations have capacity.
Then those ground stations have to connect to terrestrial fiber networks.
The computing constellation may be in space.
Its customers remain inconveniently attached to Earth.
So our orbital data center still requires enormous terrestrial infrastructure.
We have escaped the building.
We have not escaped the planet.
Now add traffic
A million satellites also create an orbital traffic problem.
Earth orbit already contains thousands of operating satellites and tens of thousands of tracked pieces of debris.
Now add a million computing spacecraft.
Their positions have to be tracked.
Their orbits have to be maintained.
Collision risks have to be managed.
Failed satellites have to be dealt with.
And every replacement satellite means an old satellite has to go somewhere.
At a five-year replacement cycle, our mature constellation potentially retires:
200,000 satellites every year.
That’s approximately:
548 satellites every day.
About:
23 every hour.
One roughly every:
Two and a half minutes.
While you’re eating breakfast, several data-center satellites have retired.
While you’re watching a football game, dozens have.
We have transformed computer replacement into a continuous orbital industrial process.
And down they come
Low-Earth-orbit satellites eventually reenter the atmosphere.
Much of a properly designed small spacecraft can burn up.
Some components can survive.
NASA has studied spacecraft reentry for decades because materials including titanium and stainless steel can sometimes survive atmospheric entry and reach the surface.
Now scale the process.
We’re no longer occasionally retiring a satellite.
Under our five-year example, we’re potentially retiring 200,000 computing satellites every year.
There are also environmental questions.
What happens when enormous quantities of spacecraft material repeatedly vaporize in the upper atmosphere?
What compounds are deposited?
What are the cumulative effects over decades?
Those questions are being studied, and we should resist the temptation to answer questions science is still working on.
Moving computing into space doesn’t necessarily eliminate environmental consequences.
It can move some of them to a place where they’re harder to see.
SpaceX deserves some credit here
SpaceX has already accomplished things that many knowledgeable people once doubted.
Routine recovery and reuse of orbital-class boosters is one of them.
Falcon 9 has fundamentally changed launch economics.
A launch cadence above 160 Falcon flights in a year is remarkable.
Starship could change the economics again if SpaceX achieves rapid and reliable full reusability.
Launch prices could fall.
Satellite manufacturing could become more automated.
Processors could become more energy efficient.
Radiators could get lighter.
Laser communications could improve enormously.
Orbital servicing could become routine.
Maybe several of Shubber’s assumptions eventually look quaint.
That’s how technological progress works.
But technological progress has one irritating limitation.
It doesn’t repeal arithmetic.
One million isn’t the interesting number
That’s what I took away from Shubber’s note.
“One million satellites” sounds futuristic.
It’s so large that the human brain doesn’t instinctively grasp it.
A million.
We’ve heard the word enough that it has become strangely harmless.
A million dollars.
A million people.
A million downloads.
A million satellites.
Fine.
But divide the million by launch capacity.
Then divide it by useful life.
Suddenly the number stops being futuristic.
It becomes logistics.
Using Shubber’s deliberately generous assumptions:
90 satellites per Starship.
500 Starship launches every year.
That gives us:
45,000 satellites per year.
It takes:
22 years to launch one million.
Except the satellites don’t last 22 years.
At a five-year replacement cycle, maintaining a completed million-satellite constellation requires approximately:
200,000 replacement satellites every year.
At 90 per launch:
More than 2,200 Starship launches every year.
More than six launches every day.
Just to keep the constellation from shrinking.
Before expansion.
Before failures.
Before weather delays.
Before launch accidents.
Before lunar missions.
Before Mars.
Before somebody else wants to use the rocket.
Before we cool a million computers.
Before we solve the ground communications bottleneck.
Before we dispose of a million spacecraft.
Before anyone asks what all this costs.
Physics doesn’t read the pitch deck
There is something genuinely attractive about putting computing in space.
The Sun provides enormous amounts of energy.
Land isn’t an issue.
Terrestrial water consumption can potentially be avoided.
Computing close to satellites makes considerable sense for certain applications.
Someday we may indeed build substantial computing infrastructure in orbit.
I think we probably will.
But there is an enormous difference between something being technically possible and being practical at industrial scale.
That’s why Shubber’s arithmetic matters.
You don’t have to decide whether Elon Musk is a genius, a huckster or some interesting combination of the two.
You don’t even have to bet against Starship.
Do the opposite.
For the sake of argument, let Starship work spectacularly.
Give it 90 computing satellites per flight.
Give SpaceX 500 Starship launches every year — roughly three times its record Falcon launch cadence.
Give every satellite five kilowatts of computing capacity.
Give the proposal every break we reasonably can.
Then take the million-satellite idea seriously.
Get out a calculator.
And start dividing.
Engineering can overcome extraordinary problems.
Physics can sometimes be worked around.
Costs can fall.
Technology can improve.
But sooner or later, somebody has to do the math.
The future can survive arithmetic.
If it can’t, it probably wasn’t the future.