Space Data Centers: SpaceX vs Google Suncatcher vs Starcloud, and Why SpaceX's Plan Looks Strongest

Google's first space data center went up on Thursday on a SpaceX rocket. A Falcon 9 lifted off from Vandenberg at 11:18 a.m. local time carrying 130 payloads for different customers, and one of them was a refrigerator-sized machine from Google's Project Suncatcher with four of Google's AI chips inside. It runs on about 1 kW of solar power, and the chips can only work for about 15 minutes at a time before they have to shut off and cool down.

SpaceX, which launched it, says each of its own AI satellites will run 175 kW on average. A third company, Starcloud, already trained a small AI model on an Nvidia GPU in orbit last year, and it wrote one of the first detailed plans for all of this back in 2024. That makes three different ideas about how to build a data center in space, and all three companies have published enough engineering to compare them.

Whichever design wins has to make two things cheap, lifting a kilogram into orbit and getting rid of the heat once it's there. Google's own cost model says orbit only gets near Earth's power bill once launch drops under about $200 a kilogram. Starcloud says it can't compete until Starship flies often, and both of them buy their rides from SpaceX. I think SpaceX gets there first because it owns the rocket and already builds satellites by the thousand.

Free sunlight, and nowhere for the heat to go

People started looking up because of the grid. The IEA expects data centers worldwide to use about 950 TWh of electricity in 2030. That's nearly double the 485 TWh they used in 2025. A data center goes up faster than the grid can hook it up, and getting connected can take years. In Texas so much new demand asked to connect that the grid operator froze 474 GW of requests, more than five times the state's record peak of about 85 GW.

Sunlight in orbit is stronger and it doesn't stop. Above the atmosphere a square meter of panel gets 1,361 watts of sunlight, against roughly 400 that reaches a panel on the ground. There's a special path called a dawn-dusk sun-synchronous orbit that rides the line between day and night all the way around the planet, so a satellite there almost never passes into Earth's shadow. Google's paper says a panel in the right orbit collects up to eight times the energy per year of the same panel at a typical latitude on Earth. So a satellite there doesn't need batteries to get through the night.

Getting rid of the heat is harder. On the ground a data center dumps it into air or water, but in a vacuum there's nothing to dump it into, so the only way out is to glow it away as infrared light from a big flat panel called a radiator. How much a radiator can shed depends on its area and on its temperature raised to the fourth power, so a hotter panel sheds a lot more than a cooler one. Andrew Cavalier at ABI Research wrote in IEEE Spectrum in June that one Nvidia H100 drawing 700 watts needs about 1.4 square meters of radiator if the radiator runs at 60°C, and close to 3 square meters at 20°C. "Every square meter of power generation now demands approximately another square meter of cooling," he wrote. Starcloud's own white paper works out to about 1,580 square meters of radiator for every MW of computing, which is about the size of an NHL hockey rink. The space station's main cooling system handles 70 kW, so 1 MW of chips would need about 14 of them.

Radiation flips bits and slowly wears chips down, and nothing up there gets repaired. Before any of that, every kilogram has to be launched.

Starcloud's 2024 white paper

Starcloud was still called Lumen Orbit when its founders, Philip Johnston, Adi Oltean and Ezra Feilden, published a white paper called "Why we should train AI in space" in September 2024. It's still worth reading because it gives a number for almost everything.

The paper pictures a 5 GW data center built from containers. Each container holds compute and gets launched whole, then docks to a central spine that carries power, networking and coolant through a single port. Solar and radiator modules bolt on around it until the whole thing is a flat sheet. The solar array alone would be about 4 kilometers on a side. One launch carries about 40 MW of computing, and the paper figures fewer than 100 launches of compute plus a similar number for panels would get you to 5 GW.

If you run a 40 MW cluster for ten years, the paper's cost table puts the bill at $167 million on Earth and $8.2 million in space. Almost all of the Earth figure is electricity, $140 million of it at 4 cents per kWh. The space side is $2 million of solar cells, $5 million for one launch and $1.2 million of radiation shielding. The chips, the structure and everything else got left out of both columns because the authors called them "approximately equivalent." The paper calls its $5 million, 100-ton launch about $30 a kilogram, though those numbers actually work out to $50. It says the price could fall as low as $10, citing a tweet from Elon Musk.

People have rerun it. In a long blog post, Angadh Nanjangud rebuilt the 40 MW case with radiators as heavy as the ones on the International Space Station and got 63,183 square meters of radiator and 17 to 22 launches instead of one. Just repricing Starcloud's one launch at $500 a kilogram turns $8.2 million into $53.2 million. His full version costs $110 million at Starcloud's own $5 million a launch, and $2.2 billion if a launch costs $100 million.

Starcloud's own numbers have moved too. In March its target energy cost was 5 cents per kWh, and only if launch gets down to $500 a kilogram. That's about what a new data center pays for grid power, where the white paper had promised two-tenths of a cent. "We're not going to be competitive on energy costs until Starship is flying frequently," Johnston told TechCrunch.

Starcloud has flown more than you'd expect from a company that size. Starcloud-1 was a 60-kilogram satellite with one Nvidia H100 that launched on a Falcon 9 in November 2025 into an orbit only 325 kilometers up. In December it trained nanoGPT (a tiny teaching version of a language model written by the AI researcher Andrej Karpathy) on the complete works of Shakespeare and ran Google's Gemma model, the first time anyone had done either with a high-end GPU in orbit. IEEE Spectrum's Dina Genkina noted the radiator on it was too weak to let the chip run at full power.

The company is now worth $2.3 billion after an August round that included $25 million from Nvidia, and it had 25 employees at the time. Its next two satellites are 8 kW each, going up on rideshares in 2027. The one after that, Starcloud-3, is a 200 kW, three-ton spacecraft designed to be stacked inside a Starship and pushed out the door the way Starlinks are. Its FCC application asks for up to 88,000 satellites. "If you want my real answer, we're not going to be docking anything for quite a long time," Johnston told TIME in January.

Starcloud also buys a lot from SpaceX. In May it signed up to fly more than 50 of SpaceX's Starlink Mini Lasers on its satellites. In August Johnston said Falcon 9 is scheduled to retire in 2028 and the other American rockets aren't flying often, and that "if we can't book any SpaceX launch capacity in 2029, that will be challenging for us."

Google's Suncatcher flies small satellites in a tight swarm

Google published its version in November 2025 as a research paper, Towards a future space-based, highly scalable AI infrastructure system design, and it goes into far more engineering detail than Starcloud's paper did. It also cites Starcloud's white paper by name as the kind of giant single structure Google decided against, because something bigger than any rocket has to be assembled in orbit by robots or people.

Google's design uses lots of small satellites instead. The example in the paper is a cluster of 81 satellites flying in formation inside a circle 2 kilometers across, at about 650 kilometers up in that dawn-dusk orbit. Neighbors sit only 100 to 200 meters apart.

They have to be that close because of the networking. Google's chips are built to work in pods, thousands of them wired together so a big model can be split across all of them, and the links between chips inside a pod run at hundreds of gigabits per second each. A laser between two satellites spreads out as it travels, so the power reaching the far end falls with the square of the distance. Google's paper puts Starlink's lasers at about 100 gigabits per second over as much as 5,400 kilometers. Google figures it needs on the order of 10 terabits per second per link, about a hundred times more. To get that out of off-the-shelf fiber-optic parts, the satellites have to fly within a few hundred kilometers of each other, and at a few hundred meters apart each link can carry several beams side by side. Google's bench test with commercial parts already pushed 1.6 terabits per second across a short gap.

Flying 81 satellites that close together sounds terrifying, and the paper spends a lot of math showing it's manageable. Left alone in simple gravity, the cluster would breathe in and out twice per orbit and come back to the same shape without burning any fuel. Earth isn't a perfect sphere, which pulls the cluster apart slowly, but Google found that stretching the formation by a fraction of a percent cuts the drift to under 3 meters per second of velocity change per year for every kilometer of cluster.

Google took a working Trillium TPU, its sixth-generation AI chip, to the cyclotron at UC Davis and fired a 67 MeV proton beam at it while it ran. With about 10 millimeters of aluminum shielding, a chip in that orbit should absorb around 750 rad over five years. The high-bandwidth memory started acting up at 2,000 rad. Everything else kept running machine-learning jobs correctly up to 15,000 rad, the highest dose they tested, and nothing failed outright from the accumulated dose. Random bit flips did turn up, about once every 3 million inferences. Google says that's probably fine for answering questions, but it's still an open question for training.

Google also fit a learning curve to SpaceX's launch prices going back to Falcon 1 and found they've fallen about 20% every time cumulative mass to orbit doubled, from at least $30,000 a kilogram on Falcon 1 to about $1,800 on Falcon Heavy. If that keeps up, launch gets under $200 a kilogram around 2035. That takes about 180 Starship launches a year to happen. At $200 a kilogram, Google estimates the cost of lifting the power hardware works out to about $810 per kW per year, and a US data center spends somewhere between $570 and $3,000 per kW per year on electricity. Neither number includes the chips, because those cost the same in orbit or on the ground.

The satellite that went up on Thursday is called MVP, and Planet Labs built it around four TPUs on one of Planet's existing satellite frames. That's roughly the compute of one server. Google originally planned to launch two satellites in 2027 and pulled the first one into 2026, accepting extra risk to get there by using a frame Planet already flies. Heat moves out through copper and aluminum to radiator panels, and the chips can run for about 15 minutes before they have to shut down to cool, Travis Beals told the New York Times. The next step is two satellites in 2027 to test the laser links.

"We don't expect, to be perfectly frank, that we'll have anything usefully operational in the next few years," senior vice president James Manyika said. Beals, who runs the project, went further: "If, five years from now, everything we've done has worked perfectly, it probably means we've not taken enough risk."

These three aren't the only ones trying. China's Three-Body Computing Constellation launched its first 12 satellites in May 2025, each running an AI model with 8 billion parameters, and it's aiming for 2,800. Blue Origin asked the FCC in March for permission to fly up to 51,600 data center satellites under the name Project Sunrise. And when Nvidia launched its space computing chips in March, it named six partners, including Starcloud, Planet, Axiom Space and Aetherflux, the startup from Robinhood's cofounder.

SpaceX's AI1 puts a whole Nvidia rack in each satellite

SpaceX's satellite is called AI1, part of a program the company calls Starmind, and it's big. The Starmind page lists it at 75 meters from wingtip to wingtip and 30 meters tall once it unfolds. The page rates its solar array at 210 kW, and on Thursday Musk raised that, posting that each satellite will have "over 250kW of solar power, which is about 20% more than the International Space Station." Its compute runs up to 250 kW at peak and 175 kW on average. Heat leaves through a 160-square-meter liquid radiator with redundant pumping loops and shielding against micrometeoroids. SpaceX hasn't published a mass, but its efficiency figure of 75 kW per ton puts one AI1 at roughly 2.3 to 3.3 tons, depending on whether that's measured at average or peak power.

The design grew this summer. When SpaceX first showed AI1 in a video before its IPO in June, it was 150 kW peak and 120 kW sustained with a 110-square-meter radiator and a 70-meter wingspan. In August Nvidia said the AI1 compute payload is a Vera Rubin NVL72, its 72-GPU rack, and Musk later said the space-optimized version launches in the fourth quarter of 2027.

AI1 has to shed 175 kW through 160 square meters, which is about 1,090 watts per square meter. The June design shed the same amount per square meter, so SpaceX grew the radiator right alongside the compute. Musk has said the panels radiate from both sides and turn knife-edge to the sun, so sunlight barely hits them. For a two-sided panel to shed 1,090 watts per square meter, it has to run at about 49°C, and at the 250 kW peak it's closer to 79°C (that's my arithmetic, using the standard radiation formula and ignoring the heat the panel soaks up from Earth). Cavalier's IEEE model works out to about 500 watts per square meter, and Starcloud's 2024 paper assumed 633. So a MW of AI1 needs about 900 square meters of radiator, a bit more than half a hockey rink. Most of that edge comes from running the panel hotter and keeping it edge-on to the sun. Cavalier's radiator only radiates from one side, and Starcloud's has sunlight falling on one face. GPUs are fine at temperatures that would burn your hand, but nobody has flown a radiator like this yet.

Google needs tight formation flying because each of its satellites holds only a handful of chips, so the cluster has to act like one big computer across a couple of kilometers of empty space. SpaceX puts a whole NVL72 in each AI1, and those 72 GPUs already talk to each other over copper inside the box. That's enough to serve a large model on its own, so the satellites only need to talk to each other and to the ground over the kind of lasers SpaceX already flies. SpaceX's IPO filing also describes linking satellites into larger clusters later, but the first version doesn't depend on it.

And SpaceX already flies a lot of those lasers. The same filing says Starlink had more than 23,000 inter-satellite lasers in orbit at the end of March, and each new V3 Starlink carries six 400-gigabit optical links. Starlink is about three-quarters of all the active satellites that can maneuver, and the factory in Redmond, Washington averaged about 70 Starlinks a week from December through April. Before AI1 flies, Gwynne Shotwell, SpaceX's president, has said the company will put compute on ordinary Starlinks first as "canary" tests.

The factory for the real thing is in Bastrop, Texas. SpaceX calls it the Gigasat Factory, 11 million square feet on more than 1,000 acres, and it's meant to make everything from the silicon for the solar cells to finished satellites. The goal is 1 GW a year of orbital compute by late 2027. At 175 kW per satellite that's about 5,700 satellites a year, or 16 a day, while the Starlink line in Redmond turns out about 10.

The first job in orbit is answering questions

Training a model takes thousands of chips trading data constantly. Using a finished model to answer questions, which the industry calls inference, needs much less back-and-forth between chips.

Starcloud's 2024 paper was called "Why we should train AI in space," but its next satellites will run inference for customers including US government agencies. SpaceX's IPO filing pitches its satellites for workloads "such as inference demand". With a whole rack in each AI1, most of the chip-to-chip traffic stays inside one satellite. Google designed for a swarm that acts like one big computer, so its satellites have to fly close together. Even so, its own radiation tests found the bit flips look fine for inference and left training as an open question.

Starship sets the price for all three space data center plans

Launch is the number all three share, both the price of a ton to orbit and how often a rocket can fly. Google's paper puts today's Falcon 9 price at about $3,600 a kilogram.

SpaceX's IPO filing says 100 GW a year of orbital compute, at about 100 kW of computing per ton of satellite, means hauling about one million metric tons to orbit every year, which takes thousands of launches a year. Starship V3 is designed to carry 100 tons fully reusable. If you fill one with AI1s at 75 kW per ton, you get about 7.5 MW of compute per flight. So 1 GW takes roughly 130 Starship launches, going by SpaceX's own two figures, and SpaceX's goal of 100 GW a year works out to about 36 launches a day. Musk told Dwarkesh Patel in February that 100 GW is "on the order of 10,000 Starship launches," and that it could be done with as few as 20 or 30 physical ships flying over and over.

Starship isn't close to that yet. It reached orbit for the first time on Monday, on Flight 14, and it was a messy success. The booster lost an engine on the way up, the ship lost one of its vacuum engines, and launch control briefly called off the orbit attempt before reversing about ten minutes later and firing a 95-meter-per-second burn to make it. The ship released all 26 of its Starlink V3 satellites, all 26 checked in with the network, and SpaceX brought it down after just under two orbits. Provisional planning documents put Flight 15 no earlier than October 19, though NASASpaceflight says the vehicles might not be ready until the end of October, and Musk has said SpaceX might try catching the ship with the tower on that flight.

Google's paper also modeled what a Starship launch costs SpaceX itself, and got about $460 a kilogram with no reuse, under $60 with each part flying ten times, and under $15 with a hundred. The propellant alone sets a floor around $8. Google then assumed SpaceX would keep a margin as high as 75%, which still puts the customer price under $250 a kilogram. Google and Starcloud would pay that price, while SpaceX launches its own satellites at cost.

What could still go wrong for SpaceX

There's a lot, and SpaceX says much of it in its own filing. "We have not, and no one else has, previously operated or attempted to operate orbital AI compute," the S-1 says, and hardware in orbit "will not be easily repaired or upgraded." It also says the satellites wear out before the computers inside them do. The FCC filing gives the system a five-year life.

The radiator claim is unproven. SpaceX needs about twice the heat rejection per square meter that IEEE's model assumes, and the only way to check it is to fly one and measure it. The cost math is also against orbit today. ABI Research says an orbital data center can cost upward of 78 times its equivalent on the ground. SemiAnalysis, which is friendlier to the idea, put the all-in cost of a top Nvidia GPU at $8.64 an hour in space against $2.37 on the ground in 2026, and its base case has the two lines crossing around 2040. Sam Altman called orbital data centers "ridiculous" at a summit in February. Jensen Huang, whose company sells the chips for it, said the same month that "the economics are poor today, but it is going to improve over time."

Space is also getting crowded. Starlink satellites made 207,152 collision-avoidance maneuvers in the six months to May, with roughly 10,000 satellites up. The million-satellite application drew about 1,000 comments at the FCC, mostly opposed, including a petition to deny from the American Astronomical Society and a petition from Amazon asking the FCC to reject it. The FCC hasn't ruled.

Why I think SpaceX figures it out

I think SpaceX gets this working first, and I got most of that from reading Google's and Starcloud's own papers.

Google's launch math is a curve fit to SpaceX's prices, so Google's plan only works if Starship flies about 180 times a year by about 2035. Google is also betting on SpaceX directly. It owns about 6% of the company, a stake worth $94.1 billion at the end of June, and it pays SpaceX $920 million a month to rent about 110,000 GPUs.

Starcloud needs SpaceX even more, since Johnston says it can't compete on power costs until Starship flies often, and in May it signed up to fly SpaceX's Starlink Mini Lasers. Its next big design is a 200 kW, three-ton satellite sized for Starship, which is about the same power and mass as AI1.

SpaceX sets its own launch price. Because it flies its own satellites at cost, each Starship flight lowers SpaceX's cost before any customer sees a lower price. The company already builds about 70 Starlinks a week and runs the largest laser network in orbit, on solar arrays it makes itself. And because a whole rack fits in one AI1, SpaceX can start selling compute from a single satellite, while Google needs a formation of 81.

The radiator is harder, and SpaceX's answer so far is to test it in orbit. Its IPO filing says many of the cooling methods it plans to use have already been proven on Starlink, and it's putting compute on ordinary Starlinks before the first AI1 goes up. Since June the design has grown from 120 to 175 kW of average compute, and the radiator grew with it, so each square meter still sheds about 1,090 watts. This week Musk added more solar on top of that. That convinces me more than any of the launch math, because a team guessing at its radiator wouldn't hold that number flat while the compute grew by half.

Starship blew up several times in testing, and then on its 14th flight it reached orbit with an engine out and still let go of all 26 satellites. SpaceX has also caught its Super Heavy booster with the launch tower on three test flights. Starlink went from about 6,000 satellites in 2024 to more than 10,000 by June. I expect AI1 to change more than once before it works. SpaceX can afford that more than the other two, because it doesn't pay anyone else for the launch. When Musk was shown an analyst's $68 billion cost estimate for 1 GW in orbit, he said it "will be way better than that by 2028."

Where the revenue lands first

Nvidia will probably book money from space data centers first, because it sells the chips. $NVDA is co-designing the AI1 compute payload, launched a line of space computing modules in March whose headline part, the Space-1 Vera Rubin Module, claims 25 times the compute of an H100, and put $25 million into Starcloud.

$SPCX books launch revenue on every Starcloud, Suncatcher and rideshare satellite in the meantime, and it plans to rent out the compute in its own satellites the way it rents out its Memphis data centers today. Shotwell said SpaceX "100 percent" sees itself as a competitor to the GPU cloud companies. $GOOGL gets paid both ways, through Suncatcher if its swarm works and through its SpaceX stake if AI1 does. Planet Labs, which builds Google's satellites, and Firefly Aerospace, which just signed Starcloud to fly a compute payload around the Moon no earlier than 2028, are the smaller public names in it.

If I'm right, the first full Nvidia rack in orbit goes up on a Starship in late 2027, inside an AI1.

Tags: spacex, google, starcloud, space data centers, orbital compute, pillar