SpaceX has asked the government for permission to launch up to a million satellites, and to use them to build a data center in orbit. That's according to the Wall Street Journal, which read the filing. The request went in as part of the company's pitch to investors, a few months before SpaceX sold shares to the public for the first time, in June 2026. Elon Musk says data centers in space are two to three years away. The head of Nvidia, the company that makes the chips all of this would run on, talks about launching them by the hundred gigawatts a year. One gigawatt is roughly what one large nuclear reactor puts out. Engineers and analysts have now checked those numbers. This article walks through what they found, one wall at a time. By the end you'll be able to tell the physics from the sales pitch.
Why would anyone put a data center in space?
The pitch starts with sunlight. Pick the right orbit and a solar panel up there never sees night and never sees a cloud. Same panel, same size, about five times the electricity it would make on a roof. And because the sun never goes down, you don't need batteries to get through the dark hours. Data centers are electricity hogs, and towns keep fighting new ones. So the argument goes: build them where the sun never sets, and beam the answers down.
That last part sounds harder than it is. A question is a few words. An answer is a few sentences. Sending those up and back down uses almost no bandwidth — the pipe barely has to be bigger than a phone call. That's why fans of the idea say you only have to beam up the bits.
The strongest version of this case comes from Will Marshall. He co-founded Planet Labs, which already flies a fleet of camera satellites, and he was a physicist at NASA before that. He says a study his company ran with Google found a specific finish line. Everything in space is priced by weight — you pay by the kilogram to get it off the ground. Today that price runs about $1,000 per kilogram, which is roughly $450 a pound. Marshall says that once it falls to $200 or $300 a kilogram, orbit becomes cheaper than the ground. Five years ago the gap was twice as wide, so the price really is falling. Marshall is all the way convinced. "It's not just plausible — it is going to happen," he says. He predicts most of the world's computing money will be spent in space within ten years. That's one practitioner's forecast, built on a study only he has seen. But he's a serious operator, and it's the best pro case on offer.
What does the math say?
A space-industry analyst who publishes his financial models so anyone can check them ran the full comparison. His finding knocks the leg out from under the pitch. Electricity is less than five cents of every dollar an AI data center costs. The chips are the money. So free sunshine in orbit saves you almost nothing, because it saves you on the cheapest ingredient. In his model, the ground data center wins every single scenario, even if launch prices reach SpaceX's dream number.
The chips also die young. An AI chip is out of date in about five years. On Earth that's fine. You pull the old chips out and slide new ones into the same building, plugged into the same power and the same cooling. In orbit, the solar wings and the cooling panels are bolted to the chips. When the chips go stale, all of it goes stale together, and staying current means launching the whole thing again.
Then there's the weight. Chip engineer Anastasi Bartel priced out the modest version that Starcloud — the startup furthest ahead — talks about: a 40-megawatt block, about a twenty-fifth of a reactor's worth of power. It comes to more than a thousand tons. And most of that weight isn't computers. It's cooling panels and solar panels. The pictures these companies release show racks of glowing chips. The actual shipping list is mostly plumbing. The rocket rides alone come to about $5 billion at today's prices.
And when engineers checked the front-runner's own homework, it didn't hold. The channel Real Engineering, working with the engineering magazine IEEE Spectrum, took Starcloud's published plan and compared it line by line against the hardware Starcloud itself names. Two numbers do the damage. Starcloud assumes it gets 400 watts of computing out of every kilogram it launches; the Nvidia racks it points to deliver 88 — less than a quarter as much. And it budgets $30 to launch each kilogram; the first known commercial Starship deal works out to about $900, which is thirty times more. Put the real numbers in and Starcloud's flagship vision turns into a station heavier than an aircraft carrier, with a launch bill past $100 billion.
What about the heat?
Here is the part most people never think about. Every computer turns the electricity it uses into heat. On Earth you blow air across it, or run water through it, and the heat goes away. Space has no air to blow. Space is a thermos. The only way to dump heat up there is to let it glow away from a big flat panel, the way a woodstove warms a room from across the room. That panel is called a radiator. The bigger the computer, the bigger the panel has to be.
For the giant designs, the panel stops making sense. Real Engineering ran Starcloud's own figures for a five-gigawatt station — five reactors' worth of computing. The radiator comes out around four kilometers tall. That's two and a half miles of panel, taller than anything humans have ever built. And to feed it, the station would have to pump an Olympic swimming pool of coolant every 40 seconds.
Shrink the idea and the answer flips. For one satellite carrying a single rack of computers, several independent engineers agree the cooling actually works on paper. Panels about the size of the ones on the space station can shed that much heat, and they'd cost roughly what cooling costs on the ground. So there are two different walls here. Heat kills the giant station. Money kills the swarm of small ones.
What else breaks?
Radiation, first. Earth's air and magnetic field block most of the particles flying around out there. Above the air, they hit your chips. One stray particle can flip a single 1 to a 0 in the middle of a calculation, and the computer never knows it got the wrong answer. The fix in use today is brute force: the HP servers already running on the space station do every calculation three times on three separate machines and take the majority vote. That works, and it also triples the power bill and triples the weight you had to launch. You can buy chips built to shrug off radiation instead, but they cost ten to a hundred times more and they run slower. And none of them are GPUs, the chip type all AI runs on. Nobody has ever built a radiation-proof GPU. Bartel designs chips for a living, and she points out the cruel part: the newest and fastest chips are also the most fragile, because their parts are the smallest. Space pushes you backward down the exact performance curve AI is racing up.
Temperature, next. A satellite in low orbit circles the Earth in about 90 minutes, so it crosses from full sun into Earth's shadow sixteen times a day. In the sun the hardware reaches roughly 120 degrees Celsius, hotter than boiling water. In the shadow it falls to 170 below. That's a 300-degree swing, sixteen times a day, for years, and metal that expands and shrinks that hard eventually cracks. Starlink survives it with blankets and small heaters — but a Starlink satellite runs on a few hundred watts of electronics. At the megawatt scale, Bartel says plainly, there is no proven solution.
Then the wiring. There are two different things an AI does, and they have very different needs. Training is how a model gets built: thousands of chips grinding on the same problem at once, shoving enormous amounts of data between each other every second. Answering is what happens when you type a question in, and it's far smaller and far simpler. Training needs those thousands of chips connected at speeds the lasers between satellites cannot come close to matching. So anything that flies can answer questions, but it can't be taught. Getting a big pile of data up there to work on is worse still. The computer architect the Journal consulted said it might genuinely be cheaper to burn the data onto a disc and launch the disc.
Even the perfect orbit has a catch, and it's a trap. To truly never see night, you have to fly high. At 500 kilometers up, where most satellites live, you still sit in Earth's shadow up to a fifth of the day on some days. Real year-round sunlight means climbing toward 1,500 kilometers. But starting around 1,000 kilometers you enter the inner Van Allen belt — a donut of charged particles that Earth's magnetic field has caught and holds in place. A satellite at 2,000 kilometers takes about a thousand times the radiation of one at 500. Endless sunshine and survivable radiation live at different altitudes, and you can't have both. SpaceX's own filing lists altitudes going all the way up to 2,000.
Failure costs different amounts in the two places, too. If the AI boom goes bust on Earth, the power plants keep making power and the chips get sold to somebody else. When a machine in orbit dies — and machines die — there is nothing left to sell. There's debris. And going up to fix it costs nearly as much as putting it there did.
And orbit itself is filling up. A 2024 study by MITRE, a research group that works for the U.S. government, put the practical carrying capacity of low Earth orbit near 100,000 satellites. Internet constellations have already claimed 35,000 to 40,000 of those slots. Here's how crowded that already feels in practice: Starlink alone reported about 300,000 collision-avoidance maneuvers in 2025 — a satellite firing its thruster to get out of something's way, roughly 800 times a day. Against a ceiling of 100,000, SpaceX has filed for up to a million. Marshall, who spent years at NASA researching space debris, is more relaxed. Near-Earth space has on the order of a thousand times the room the surface of the planet has, he argues, and what's really missing is traffic rules. Room measured by volume and room measured by coordination are two different questions, and that's a genuine split between experts, not a settled matter. There's an environmental question forming too. Dead satellites burn up on the way down, and Starlink already drops about two a day. Scientists are studying the layer of aluminum-oxide dust those burn-ups leave in the upper atmosphere, and what it might mean for the ozone layer.
What a real orbital data center would take
A veteran technologist named Richard Campbell built one on paper at a conference, using only parts that already exist. Start with a single top-end Nvidia rack, the kind that can run the biggest AI models made. Add solar wings the size of the space station's. Add two cooling panels the size of the space station's. Fly it in an orbit that never sees night, so it needs no batteries at all. The result weighs 12 to 15 tons and stretches up to 150 meters across — about a football field and a half. That's bigger than any commercial satellite ever flown, though not crazier than spy satellites already up there. His verdict on that one satellite: "the math more or less makes sense."
Then he added the part the pretty pictures never show. In a data center on the ground, thousands of racks are lashed together with cables that carry data hundreds of times faster than the best laser link between two satellites. That wiring is the whole trick — it's what lets thousands of separate chips behave like one enormous chip. Nothing in orbit comes close to it. So you could fly a thousand of Campbell's satellites and still not have what one building in Virginia has. You'd have a thousand separate computers, not one big one.
Is there any technology that changes the math?
One candidate, from IEEE Spectrum's reporting: liquid-droplet radiators. Instead of pumping coolant through a giant panel, you spray the hot coolant straight out into space as a glowing mist. Each droplet dumps its own heat directly. A catcher on the far side collects them, cooled, and sends them around again. No giant panel at all. Real research, genuinely promising, and nowhere near flying.
Google has an entry too, called Suncatcher. It fixes several of Starcloud's mistakes by flying 81 small satellites in tight formation instead of one huge machine. The catch turned up in Google's own published paper. The 81 have to hold that formation precisely, so when one of them has to dodge a piece of debris, all 81 may have to move with it. Given how often satellites are already dodging up there, that is not a rare event to design around.
Meanwhile the talk keeps outrunning the machinery. Nvidia's Jensen Huang has said he expects to launch orbital data centers with SpaceX at "100 to 200 gigawatts per year," and to go to the moon for more. Set that against the arithmetic. One gigawatt of orbital computing works out to roughly 166 Starship launches. The low end of his range is a hundred times that, every year — about 17,000 launches, or 45 every single day. Quote the man exactly, and let the math do the grading.
So why is everyone suddenly talking about it?
Three separate sources looked at the timing and landed on the same reading. In February 2026, SpaceX absorbed Musk's struggling AI company, xAI. SpaceX's first official plans for space data centers appeared in the same stretch. Then came the filing to sell shares to the public — from a founder who had spent years saying SpaceX would never go public before there was a city on Mars. Campbell walked through that chain step by step at the same conference, and he was careful to call it his interpretation rather than a fact. The logic is simple. A company that owns a trillion-dollar orbital-AI story is worth more per share than one that doesn't. And the filing itself concedes, pages down, that limits on how many satellites SpaceX is allowed to launch could affect what the offering is worth.
That's their read, not a proven motive. What is documented is the order of events: merger, announcement, filing, all within months.
Who would actually buy this?
Watch what the front-runner does instead of what it says. The analyst with the published model reports that Starcloud has already made defense and surveillance computing its main business. That means taking the pictures spy satellites collect and processing them right there in orbit, then selling that service at military prices — which is the one column where his spreadsheet finally turns a profit. Real Engineering reached the same conclusion on its own. And it solves a real problem. Terabytes of classified imagery are already sitting up there, stuck behind a radio link no wider than a garden hose. Doing the work in orbit instead of shipping the raw pictures down is genuinely useful, and the customer has famously deep pockets. Huang has framed processing sensor data on orbit as essentially the reason to do any of this at all.
So the first real customer isn't your chatbot. It's the Pentagon. Which means the money that would pay for it is the defense budget, and the defense budget is voted on, by name.
Will it ever happen?
Start with the moon versions now being floated, because physics answers those fast. Light takes 1.3 seconds to cross from the moon to Earth, so every question you ask waits 2.6 seconds for its answer to come back. Nobody is going to chat with a lunar AI. Night on the moon lasts 14 Earth days, which means solar power stops for two weeks at a stretch. And landing something on the moon costs about twenty times what putting it in low orbit costs. What survives is the strangest niche in the pile: cold storage — a backup copy of humanity's most important records, parked somewhere no war or fire can reach. One startup has already tried it. Its first test payload weighed one kilogram, and it crashed on landing in March 2025.
The fairest one-line verdict in the whole pile comes from Bartel: "Physics says it's possible. Economics says not yet." She also flags why SpaceX alone stays so bullish. It is the only player that pays itself wholesale for the rockets everybody else buys retail. The same computer architect the Journal consulted put the state of knowledge honestly. The costs are large "and at the moment unknown," because nobody has flown the experiments that would settle it. One more wrinkle, from Marshall: a couple of Chinese companies, he says, have been at this longer than any American one.
That's the whole picture. The direct links, the filings, and the source data behind every number are listed below.
Where this article's evidence comes from
Where a sentence above rests on one of these sources, it links straight to it. Here is each source in one place. Open a link, check the claim against the original, and make your own call.
- The Wall Street Journal — "SpaceX Wants to Blast Data Centers Into Orbit. Here's What It May Take." — the FCC application for up to a million satellites as part of the investor pitch; the computer architect on launching a disc of data, and costs that are large "and at the moment unknown."
- SpaceNews — "SpaceX files plans for million-satellite orbital data center constellation" — SpaceX asked the FCC to approve a constellation of up to one million satellites meant to work as an orbital data center.
- The Next Web — SpaceX's IPO filing versus Musk's pitch — Musk said orbit would be the cheapest place for AI "within two years, maybe three at the latest"; SpaceX's own IPO filing warns the plan "may not achieve commercial viability."
- Council on Foreign Relations, The Spillover — Planet Labs co-founder Will Marshall — the Google-Planet tipping point at $200-$300 a kilogram, about $1,000 a kilogram today, five times the power per panel, his ten-year forecast, the room in orbit, and the Chinese companies.
- Noise In Space — "The Truth About SpaceX's 'Orbital Datacenters'" — the published cost model (spreadsheet linked from the video page): electricity under five cents of the dollar, five-year chip life, the sunlight-versus-radiation altitude trade, 166 Starship launches per gigawatt, the xAI timing, Starcloud's defense turn, and Huang's sensor-processing framing.
- Real Engineering, with IEEE Spectrum — Starcloud's white paper checked line by line — 400 versus 88 watts per kilogram, $30 versus about $900 a kilogram, the $100 billion launch bill, the four-kilometer radiator and the pool of coolant every 40 seconds, the space-station triple computing, 300,000 Starlink dodges in 2025, Suncatcher's 81 satellites moving together, and the military customer.
- Anastasi In Tech, with IEEE Spectrum — chip engineer Anastasi Bartel — the 40-megawatt block at over a thousand tons and about $5 billion to launch, radiation-hardened chips at ten to a hundred times the cost, the 300-degree swing, aluminum-oxide from burn-ups, the moon's 2.6-second delay and 14-day night, the crashed one-kilogram test, "Physics says it's possible. Economics says not yet." The video also carries Jensen Huang's "100 to 200 gigawatts per year" remark.
- NDC Copenhagen 2026 — Richard Campbell, "Above the Cloud: Building Data Centers in Space" — the one-rack satellite (12 to 15 tons, up to 150 meters), ground cables versus laser links, the 2024 MITRE carrying-capacity estimate near 100,000 satellites, 35,000 to 40,000 already claimed, and the share-offering walk-through he labels as interpretation.
- IEEE Spectrum — Andrew Cavalier on cooling in orbit — the liquid-droplet radiator idea: spray the coolant into space and catch it again.
- Google Research — Project Suncatcher — Google's own design and published paper: an 81-satellite cluster flying hundreds of meters apart.
- CNBC — Starcloud trains the first AI model in space — background on the front-runner: Starcloud-1 put an Nvidia H100 chip in orbit in November 2025.
- SpaceNews — China launches the first of 2,800 computing satellites — a second account behind Marshall's China point: ADA Space and Zhejiang Lab launched the first 12 satellites of an orbital computing network in May 2025.
- Jensen Huang, Nvidia — the "100 to 200 gigawatts per year" and moon remarks, and the sensor-processing framing, quoted from the clip itself (2026).