Data Centers in Orbit: SpaceX's 100 GW Plan Needs 10,000 Starship Launches
SpaceX's prospectus wants 100 gigawatts of data centers in orbit every year. Musk's own math says that takes about 10,000 Starship launches, and Starship has flown thirteen times. What the plan actually needs, in radiators and rockets, and who gets paid before any of it works.
Musk says space will be the cheapest place to run AI within three years. The prospectus says nobody has ever tried. Whether he's right comes down to how often Starship flies.
Disclosure: Impersonal research, not personal investment advice (see the disclaimer). Figures last refreshed 2026-08-23; sources linked inline. Evergreen piece, refreshed quarterly.
Starship came back from its latest test flight in late July so gently that it floated in the Indian Ocean. Twenty Starlink satellites rode up with it and burned up on purpose about twenty minutes later, which was the plan. Somewhere in SpaceX's IPO paperwork, that same flight is the first step toward a million computers in the sky.
That's the pitch, anyway. Elon Musk told the Davos crowd in January that space would be the lowest-cost place to run AI "within two years, maybe three at the latest" and called data centers in orbit a no-brainer. A few weeks later he told Dwarkesh that 100 gigawatts of compute in orbit takes "on the order of 10,000 Starship launches." Then the lawyers wrote the prospectus, and the prospectus says the orbital plan rests on "unproven technologies, or technologies that do not exist."
A data center in orbit has two hard problems, heat and launch, and SpaceX only controls one of them. How often Starship flies decides the rest. So most of this is about the launch count, because that's the part that moves.
The setup: a no-brainer with a risk factor
A year ago orbital compute was a startup pitch. Starcloud (a small company whose satellite was built by Astro Digital) put the first Nvidia H100 into orbit in November 2025 and trained a toy language model on Shakespeare from 325 kilometers up. Google announced Project Suncatcher, a research project to fly its own TPU chips on solar satellites, with two prototypes going up with Planet Labs by early 2027. Small, careful science.
Then SpaceX bought xAI in February, filed with the FCC for up to one million satellites between 500 and 2,000 kilometers up, and made orbital AI the headline of the largest IPO ever. The S-1 wants 100 gigawatts of compute per year carried to orbit, starting as early as 2028. It says that takes "thousands of launches per year" and "approximately one million metric tons of payload annually." The whole planet put about 2,213 metric tons into orbit in 2025. (So, a 450x increase, from a company that still counts Starship flights on its fingers and toes.)
The same document also says this, word for word: "We have not, and no one else has, previously operated or attempted to operate orbital AI compute." And this: the timeline for 100 gigawatts "may be difficult or impossible to determine." I read a lot of risk factors. Most of them are boilerplate about hurricanes. This one reads like an engineer got hold of the pen.
The stock has already voted once. SpaceX's AI segment lost $1.3 billion in the second quarter and the shares sit under the $135 IPO price. That loss belongs to the ground business, Colossus in Memphis, the one that exists. The orbital business doesn't have a revenue line yet.
What a data center in orbit actually needs
Musk is right about the sun. A solar panel in a dawn-dusk sun-synchronous orbit (a path that rides the line between day and night, so the panel faces the sun almost all the time) never sees a cloud. Musk says those panels run "about five times harder" than panels on the ground, with no batteries. SemiAnalysis, which has done the most careful public math on this, puts the time in Earth's shadow at up to 35 minutes a day in that orbit. So power is fine.
Heat isn't. On Earth a data center dumps its heat into air or water. In a vacuum there is neither, so the only way to lose heat is to glow it away from a big flat panel. Andrew Cavalier at ABI Research worked the numbers for IEEE Spectrum in June. One 700-watt H100 needs about 1.4 square meters of radiator. A single 40-kilowatt rack needs 80 square meters, roughly a pickleball court. The panels also fade, so after five years you need 40% more area for the same chips. His line on it: "In orbit, the only variable we can control is area."
SpaceX's own filing agrees with the physics. It says shedding one megawatt of heat while keeping the electronics at 20 degrees takes "approximately 1,200 square meters, roughly the size of four tennis courts" of radiator. The first satellite, which SpaceX calls AI1 and showed off in June, is 70 meters tip to tip and 20 meters tall. It runs 120 kilowatts of compute on Nvidia chips, with 110 square meters of radiator folded out behind it at 600 kilometers. Musk called it "much simpler than a Starlink satellite," and production is supposed to start at the Bastrop, Texas factory by the end of 2027.
Here's where I got stuck. Run the SpaceX drawing through the IEEE model and AI1 sheds about 1,100 watts per square meter of radiator. The independent model allows about 500. (That's my arithmetic, 120 kilowatts over 110 square meters against 40 kilowatts over 80.) So either SpaceX runs its chips a lot hotter than 60 degrees, or the drawing is optimistic by about half. For scale, the radiators on the International Space Station reject about 70 kilowatts across 325 square meters and cost somewhere between $340 million and $500 million. One AI1 has to beat that by a wide margin, at a price that still makes sense times a million.
Then radiation. Google's testing found its TPUs can survive a five-year mission with shielding, and nothing gets repaired up there. SemiAnalysis models 20% spare hardware in orbit against 5% on the ground, just to keep a cluster whole.
Then the link home. A satellite in low orbit passes over any one ground station for five to seven minutes a day. Everything else routes satellite to satellite, which adds 30 to 80 milliseconds each way. Fine for training a model. Annoying for a chatbot.
Add it up and SemiAnalysis gets $8.64 an hour per GPU in orbit against $2.37 on the ground, all in. Their base case has the two lines crossing around 2040. Their Musk case, where Starship flies the 10,000 times he says it will, gets orbit to within 30% of Earth in the early 2030s. Launch is the biggest single line in that model: $1.6 million of a $3.1 million reference satellite is the ride up. Cut the launch cost and the rest of the plan starts to pencil.
The map: three orbital data center programs, one rocket
Three programs, and all three need the same truck. Google is in talks with SpaceX to launch Suncatcher. Starcloud-1 rode up on a SpaceX rocket. Whoever wins the race to compute in orbit, the hauling is SpaceX's either way. (The rocket company gets paid whether the satellites compute or not.)
Who gets paid before any of it works
Orbit pays three public companies, in a particular order, and SpaceX is last.
Nvidia gets paid first. AI1 launches on Nvidia chips, and the S-1 admits SpaceX needs "significantly more than are currently available to us." A GPU sold to SpaceX for a satellite shows up on $NVDA's data center revenue line the same quarter as a GPU sold to Microsoft for Iowa. The invoice doesn't care about the address. The chips have to survive radiation for the orders to keep coming, but the first few thousand ship before anyone knows.
Rocket Lab gets paid second. It already has the largest production capacity in the world for the gallium arsenide solar cells satellites use today, and in February it added a silicon array line built for gigawatt-class orbital data centers. Peter Beck) called space data centers "the next frontier in computing infrastructure," which is what CEOs say, but the product underneath is specific: a cheaper, lighter panel you can make by the acre. A 70-meter wing on a million satellites is a lot of acres. Rocket Lab books that as hardware orders in its Space Systems segment. (It also took a $23.9 million CHIPS Act award in October to expand the Albuquerque cell plant, which is the US government pre-paying for the same capacity.)
SpaceX gets paid last, and only in its own accounts. Revenue from compute in orbit doesn't exist until a customer rents a GPU in the sky, and the prospectus puts that at 2028 at the earliest. Until then the AI segment's income is Colossus rent from Anthropic and Google on the ground (see our earlier coverage: The Anthropic IPO Arrives With a $15 Billion Rent Bill). The IPO priced at 94 times sales because of orbit. The sales come from Memphis.
What I'm watching
Three things, in the order they happen.
Flight 14, now no earlier than mid-September after Musk had said end of August on the earnings call. Starship puts Starlink V3 satellites into a working orbit for the first time. The ship catch got pushed too; Musk says they will try the tower catch "in a few months." So this flight proves the orbit, and the catch waits. Then the count. Morningstar's back-of-envelope says the S-1 needs about 6,667 flights a year. I'd guess the stock trades on that flight log for the next year.
Early 2027, Suncatcher's two prototypes. If Google's TPUs come through a year of radiation on a Planet Labs bus, then AI chips can live in orbit and everyone knows it. The launch contract also tells you whether Google pays SpaceX to build a competitor's constellation.
2027, the first AI1 units. SpaceX's roadmap has two prototypes going up early in the year; Gwynne Shotwell has said late 2027, so take the later one. I want one number out of them: the measured radiator area per kilowatt after a few months in the sun. SpaceX's drawing says 110 square meters, and the IEEE math says that's too small.
The view
Space will probably be the cheapest place to run AI someday. The company doesn't know when, the founder says three years, and either way it comes down to how often Starship flies.
Musk's number is 10,000 launches. Starship has flown thirteen times.