Orbital Data Centers

Orbital data centers, also called space-based data centers, are satellites designed to run large-scale computing in orbit rather than on the ground. The idea spans everything from shoebox-sized processors that crunch another satellite's imagery to proposed constellations of thousands of spacecraft training artificial intelligence models on solar power. Advocates argue that orbit offers what terrestrial data centers increasingly struggle to secure: abundant energy without grids, land, cooling water, or permits. Skeptics answer that launch costs, heat rejection, radiation, and the lack of any repair option make orbit an expensive place to run a computer.

Between 2024 and 2026 the concept moved from paper studies to flown hardware. A European Commission study concluded in 2024 that space data centers are technically feasible.^[1] The startup Starcloud operated the first data-center-class GPU in orbit in late 2025,^[2] Axiom Space launched the first commercial free-flying compute nodes in January 2026,^[3] and Google is building prototype satellites around its AI accelerator chips.^[4] The most ambitious proposal came from SpaceX, which asked US regulators in January 2026 for permission to fly up to one million computing satellites.^[5]

Why orbit#

The case rests on sunlight. A satellite in a dawn-dusk sun-synchronous orbit rides the day-night boundary and stays in sunshine almost continuously, so it needs little battery storage; Google calculates that a solar panel in such an orbit can produce up to eight times more energy per year than the same panel on the ground.^[4][6] SpaceX's regulatory filing makes the same argument in commercial terms, claiming that satellites "directly harnessing near-constant solar power with little operating or maintenance cost" will achieve what it calls "transformative cost and energy efficiency," and that within a few years the cheapest place to generate AI compute will be space.^[5]

The other arguments concern what data centers consume on the ground. The European ASCEND study, funded by the European Commission and coordinated by Thales Alenia Space, framed orbital computing as a way to expand European capacity with no demands on land or cooling water. Its 2024 results validated the concept technically but attached a condition: the climate case works only with a launcher roughly ten times less emissive over its lifecycle than current rockets. The study's target is one gigawatt in orbit before 2050.^[1]

The cooling problem#

Space is often described as cold, but cooling there is not free. A vacuum removes convection: with no air or water to carry heat away, a spacecraft can shed waste heat only by radiating it as infrared light, and the Stefan-Boltzmann law ties the radiated power to the radiator's area and its temperature raised to the fourth power.^[7] Because chips must be kept near ordinary operating temperatures, the required areas get large. One analysis calculated that a single NVIDIA H100 GPU drawing 700 watts needs about 1.4 square meters of radiator, a 40-kilowatt rack of them roughly 80 square meters, and that ultraviolet light and atomic oxygen degrade radiator surfaces enough to demand about 40 percent more area after five years in orbit; it also found that solar arrays and radiators together account for 65 to 70 percent of a computing satellite's mass.^[7] A 2026 preprint analysis reached similar numbers, putting the radiator area for a one-megawatt orbital data center at about 2,500 square meters.^[8] Google's own feasibility paper lists thermal management among the problems still unsolved.^[6]

Radiation and off-the-shelf hardware#

Space computing has traditionally meant radiation-hardened processors far less capable than current commercial chips; orbital data centers only make sense if ordinary hardware can survive. When Hewlett Packard Enterprise sent an unmodified commercial supercomputer, Spaceborne Computer-1, to the International Space Station in 2017, some estimates gave it days before radiation-induced errors killed it; it ran for more than 18 months, protected by software that throttled operations whenever behavior drifted out of bounds rather than by physical hardening.^[9] Google reached a similar conclusion from ground testing: its Trillium TPUs, fired with a 67 MeV proton beam, showed memory irregularities only after roughly three times the radiation dose a shielded satellite would absorb in five years, with no permanent failures at the highest doses tested.^[6] Starcloud-1 carries its H100, with 80 gigabytes of memory, in low Earth orbit.^[2]

Launch costs, bandwidth, and the business case#

The binding constraint is the cost of putting mass in orbit. A Falcon 9 launch runs about $7,000 per kilogram to low Earth orbit, while SpaceX's aspirational target for Starship is $10 per kilogram; at current prices, The Register estimated, launching a single Starmind satellite would cost more than $23 million.^[10] Google projects that if launch prices keep falling to below about $200 per kilogram by the mid-2030s, a space data center could cost about the same per kilowatt-year as the energy bill of a terrestrial one.^[6] The 2026 preprint is tougher: for a general-purpose system serving customers on Earth, it finds the combined budget for building and launching hardware must fall to $250 to $1,000 per kilogram, several times below today's Falcon 9 benchmark, and closes only for systems kept busy over long lifetimes.^[8] Both analyses depend on fully reusable rockets delivering prices no vehicle has yet demonstrated.^[7]

Moving data is the other bottleneck. Optical links between satellites are fast, and Google has demonstrated 1.6 terabits per second on the bench,^[6] but capacity to the ground is limited and laser downlinks stop working under cloud. The analyses converge on the same near-term market: processing Earth-observation imagery and other satellite data in orbit, where sending finished results instead of raw data saves scarce downlink bandwidth.^[7][8]

Early demonstrations#

The precursors were experiments in space-based edge computing. After Spaceborne Computer-1's 2017-2019 run,^[9] HPE launched Spaceborne Computer-2 to the station on a Northrop Grumman resupply flight in February 2021, shifting from proving survival to doing useful work.^[11] The system has processed astronaut DNA sequencing data on board, reducing 1.8 gigabytes of raw data to a 92-kilobyte result in minutes instead of waiting on downlink, and it returned to the station with upgraded storage in 2024.^[12]

Lonestar Data Holdings pushed storage beyond Earth orbit entirely, selling backup copies of critical data on the Moon as disaster recovery beyond the reach of war or natural hazards. After relaying a copy of the Declaration of Independence through a lunar lander in 2024, the company flew Freedom, a 1-kilogram, 8-terabyte solid-state payload, on Intuitive Machines' IM-2 lander Athena, launched February 26, 2025 on a Falcon 9.^[13] Athena reached the lunar surface on March 6, 2025 but tipped onto its side in a crater near the south pole, and the mission ended days later when its batteries drained.^[14]

Programs and proposals#

Axiom Space took the incremental route: a shoebox-sized prototype called AxDCU-1, running Red Hat edge software, went to the space station in 2025, followed on January 11, 2026 by two free-flying Orbital Data Center nodes hosted on Kepler Communications satellites. The nodes carry 2.5-gigabit-per-second optical links, sell processing and storage to other spacecraft, and are meant to grow into a network measured in megawatts.^[3]

Google announced Project Suncatcher on November 4, 2025 as a research moonshot rather than a product: TPU-equipped satellites connected by free-space optical links, sketched in a preprint as clusters of 81 spacecraft about a kilometer across with neighbors 100 to 200 meters apart.^[4][15] Planet will build and operate two prototype satellites, targeted for launch by early 2027, to test the chips and links in orbit.^[4]

Starcloud, an NVIDIA-backed startup formerly named Lumen Orbit, launched the 60-kilogram Starcloud-1 on a Falcon 9 rideshare on November 2, 2025, carrying the first H100 into orbit.^[2] In December 2025 it trained NanoGPT, a small language model created by Andrej Karpathy, on the complete works of Shakespeare, the first AI model trained in space, and ran Google's Gemma model in orbit.^[16] The company has applied for a constellation of up to 88,000 three-tonne, 200-kilowatt satellites, about 20 gigawatts of compute, with the larger Starcloud-2 planned to open commercial service in sun-synchronous orbit around 2027.^[17]

NVIDIA moved from passenger to platform vendor at its GTC conference on March 16, 2026, announcing the Vera Rubin Space-1 module, a space-rated version of its Vera Rubin computer that it says delivers up to 25 times the AI compute of an H100 for in-space inference; Axiom Space, Starcloud, Planet, and Kepler are among the named partners.^[18]

Starmind is SpaceX's entry, and the largest by far. The company's January 2026 filing with the Federal Communications Commission describes up to one million satellites between 500 and 2,000 kilometers, in orbits chosen to maximize sunlight, acting as distributed AI processing nodes linked by laser; the FCC had not approved full deployment as of August 2026.^[5] On August 4, 2026, SpaceX and NVIDIA announced they would jointly develop the compute payload for the first satellite, Starmind AI1: a spacecraft about 30 meters tall with a 75-meter solar wingspan and a 250-kilowatt payload, roughly one Vera Rubin NVL72 rack, with prototypes targeted for 2027.^[10][19] SpaceX said it will use NVIDIA GPUs exclusively, with Elon Musk calling Vera Rubin "the best AI computer."^[19] Fully built, the constellation would be far larger than Starlink, currently the largest satellite network in orbit.^[5]

Criticism#

Named skepticism has kept pace with the announcements. Gartner analyst Bill Ray described the idea that orbiting servers can replace terrestrial cloud infrastructure as "peak insanity," citing temperature swings between roughly 100 and 400 kelvin, space-rated solar panels costing around 1,000 times their terrestrial equivalents, downlinks interrupted by cloud cover, and the absence of any way to send a repair technician; Gartner expects orbital computing to serve space-based customers, not Earth's cloud.^[20] An ABI Research analysis reached a similar verdict on cost, estimating that launching and running a GPU in space for a year is at least an order of magnitude more expensive than on the ground.^[7]

Astronomers object to the scale. Modeling by researchers Samantha Lawler, Aaron Boley, and Hanno Rein found that under constellations already filed, one in 15 visible points in the night sky would be a satellite, and that adding a million data-center satellites would put more visible satellites than naked-eye stars over large parts of the sky for much of the night. They also note that SpaceX's filing omits satellite dimensions, exact orbits, and casualty analysis for deorbiting spacecraft, and that satellite crowding compounds both collision risk and the existing space debris problem.^[21]

References#

1. Thales Alenia Space reveals results of ASCEND feasibility study on space data centers - Thales Alenia Space.
2. Nvidia Sends a Powerful GPU to Space - IEEE Spectrum.
3. Orbital Data Centers - Axiom Space.
4. Meet Project Suncatcher, a research moonshot to scale machine learning compute in space - Google.
5. SpaceX files plans for million-satellite orbital data center constellation - SpaceNews.
6. Exploring a space-based, scalable AI infrastructure system design - Google Research.
7. Why Thermodynamics Rules Future Orbital Data Centers - IEEE Spectrum.
8. Orbital Data Centers: Spacecraft Constraints and Economic Viability - arXiv (preprint).
9. Cutting-Edge Computing Goes Spaceborne - NASA Spinoff.
10. Elon pledges to give Nvidia a virtual monopoly over the stars - The Register.
11. Hewlett Packard Enterprise Returning to the ISS - ISS National Laboratory.
12. HPE Spaceborne Computer-2 returns to the International Space Station - Hewlett Packard Enterprise.
13. Is It Lunacy to Put a Data Center on the Moon? - IEEE Spectrum.
14. Private Intuitive Machines moon lander declared dead after falling on its side in crater at the lunar south pole - Space.com.
15. Towards a future space-based, highly scalable AI infrastructure system design - arXiv (preprint).
16. Nvidia-backed Starcloud trains first AI model in space as orbital data center race heats up - CNBC.
17. Starcloud's path to 88,000 computing satellites - SpaceNews.
18. NVIDIA Launches Space Computing, Rocketing AI Into Orbit - NVIDIA.
19. SpaceX earnings takeaways: Soaring AI costs outweigh revenue beat in first report since IPO - CNBC.
20. Orbital datacenters are a pie-in-the-sky idea: Gartner - The Register.
21. A million new SpaceX satellites will destroy the night sky - Phys.org (The Conversation).