Why put a data center in space? What has flown, and why heat is the hard part
Space offers endless solar power and no land or water use — but no air to carry heat away. What has actually flown, what is only announced, and why cooling sets the limit.
What has actually flown
The honest count, as of November 2026, is small: a few single satellites and a handful of small nodes, each carrying a fraction of one terrestrial rack. Nothing in orbit today is a data center in the sense the word is used on Earth — no facility, no campus, no paying customers at scale. What has flown is worth listing precisely, because the gap between it and the announcements is the story.
Starcloud-1 — REAL. The furthest along of the AI-focused attempts. A 60 kg satellite carrying one NVIDIA H100, launched November 2, 2025, on a SpaceX Falcon 9 rideshare (Bandwagon-4) into an orbit about 500 km up. Starcloud says the satellite ran Google's Gemma model and trained a small teaching model (nanoGPT) on the works of Shakespeare in orbit — reported by CNBC in December 2025, not independently verified. A second chip aboard, an NVIDIA A6000, failed during launch. Starcloud-2 — two satellites at 8 kW each — has slipped from October 2026 to 2027, and Starcloud-3 is a 200 kW design with no firm date. Starcloud applied to the FCC in February 2026 for authority to operate up to 88,000 satellites; an application is not an approval.
Google Project Suncatcher — REAL (prototype). Announced November 4, 2025, as a Google Research moonshot: compact constellations of solar-powered satellites carrying Google's TPUs, linked by laser beams. The first prototype, built with Planet and nicknamed MVP, launched October 1, 2026, on SpaceX's Transporter-18 rideshare. It carries four Trillium TPUs and about 1 kW of solar power — roughly one terrestrial server's worth of compute, running in bursts of about 15 minutes at a time because the heat has nowhere to go. On October 30, 2026, Google said the satellite returned its first AI response from orbit: a Gemma 4 model answering a prompt sent up from the ground. The next step is two satellites in 2027 to test the laser links between them. Google is careful to say this is a hardware test, not a data center.
Axiom Space — REAL. The most operational of the group, and the least like an AI data center. Its AxDCU-1 prototype ran on the International Space Station in fall 2025, and its first two free-flying orbital data center nodes launched January 11, 2026, riding Kepler Communications' optical relay satellites, with 2.5 gigabit-per-second laser links. In October 2026 Axiom confirmed the two nodes could talk to each other and to the ground. These are connected edge-computing nodes for spacecraft and government customers, not AI training hardware.
China's Three-Body Computing Constellation — REAL. The largest flown system, and the only laser-linked computing constellation in orbit. Twelve computing satellites launched May 14, 2025, on a Long March 2D rocket, built by ADA Space with Zhejiang Lab — about 5 peta-operations per second combined, with the fastest single satellite rated at 744 tera-operations per second, 30 terabytes of storage, and laser links between satellites at up to 100 gigabits per second. Two more satellites followed in July 2026. The stated plan is 2,800 satellites. The workloads so far are edge computing and AI models run in orbit, not a commercial data-center service.
Lonestar "Freedom" — REAL (partial). The first dedicated data-center hardware on another world. A roughly 2 kg storage unit — 8 terabytes on Phison SSDs, a Microchip FPGA, running Ubuntu — flew to the Moon aboard Intuitive Machines' IM-2 Athena lander, launched February 26, 2025. The lander touched down March 6, 2025, near the Moon's south pole, tipped onto its side, and lost power the next day. The payload powered on and achieved partial validation before the power ran out. It was a storage test, not compute, and it is now dead on the lunar surface. Lonestar's earlier Independence payload had already tested data storage from the Moon's surface in 2024, aboard Intuitive Machines' IM-1 lander — but as a software test, not a dedicated hardware unit.
Everything else is ANNOUNCED or SPECULATIVE. SpaceX filed with the FCC in January 2026 for up to one million orbital data-center satellites; its own IPO filing warns investors that orbital data centers "involve significant technical complexity and unproven technologies, and may not achieve commercial viability." The European ASCEND study, led by Thales Alenia Space and funded by the European Commission, concluded in June 2024 that space data centers could be technically, economically and environmentally feasible — but only with a launcher ten times less emissive over its lifecycle than today's, and its own roadmap puts a 50 kW proof-of-concept around 2031 and 1 GW before 2050.
Why space looks attractive
The pitch has three parts, and all three are real — they are just not the whole story.
Sunlight. In the right orbit, a solar panel never sees night. Google's design flies in a dawn-dusk sun-synchronous orbit, riding the line between day and night so the panels stay in near-constant sunlight; its paper says a panel there collects up to eight times more solar energy per year than the same panel at mid-latitude on Earth. There is no atmosphere to dim the light, no clouds, and no need for the giant batteries a terrestrial solar farm needs to survive the night.
No land, no water, no neighbours. A satellite needs no land purchase, no water permit — see how much water a data center uses — no grid connection — how data centers connect to the power grid on Earth is itself a multi-year project — and no community meeting. The U.S. Government Accountability Office, in its April 2026 assessment, said space-based data centers could reduce the land, electricity and water that data centers need on Earth — which is the strongest version of the case.
Data born in orbit. Satellites already generate more data than they can send down. Processing imagery or sensor data where it is collected — on the satellite that collected it — means sending home answers instead of raw data. The GAO says this, processing space-generated data, is the application closest to maturity. It is a much smaller ambition than replacing terrestrial AI training, and a much more credible one.
The heat problem
This is the part the pitch usually skips, and it is the reason the video this page accompanies is called what it is. Space is not cold in any sense a thermal engineer can use. It is a vacuum, and a vacuum is not cold — it is empty.
On Earth, a data center dumps heat into air or water and moves that fluid away. Convection — a fluid physically carrying heat off and replacing itself with something cooler — does the heavy lifting at every stage, from the fan on a chip to the cooling tower outside the building. In orbit there is no fluid to move to the outside world. The only way a spacecraft can finally get rid of heat is to turn it into infrared light and radiate it into the sky. Radiation is, by a wide margin, the least effective of the three ways heat moves, and in space it is the only one you get.
Follow a single watt. It starts in the silicon of a GPU, whose junction runs at 70 to 85 °C. A cold plate pressed against the chip picks the heat up into a liquid coolant — that part still works in space. The coolant is pumped through the spacecraft to a radiator, because without gravity or air you cannot rely on natural circulation. Only at the radiator does the heat finally leave, as infrared photons streaming into space. The orbital machine has to build, launch and power an entire fluid system just to carry heat to the one surface that can shed it.
How much radiator you need follows from one equation, the Stefan–Boltzmann law: the power a surface radiates rises with the fourth power of its absolute temperature. Run it for a radiator held at a realistic electronics-cooling temperature — 40 to 80 °C, with the high-emissivity coatings real spacecraft use — and a square metre of radiator sheds only a few hundred watts. A useful rule of thumb: rejecting 1 kW of heat takes about 2.5 square metres of radiator. The International Space Station, the most mature benchmark, rejects 70 to 100 kW of waste heat through 422 square metres of radiators — about 166 to 237 watts per square metre in practice, well below the theoretical maximum, because the sun and the Earth both shine heat back at it. In low orbit the sky is not cold either: the spacecraft hangs a few hundred kilometres above a warm planet that fills much of its sky with infrared glow.
Scale that up and the problem becomes the design. A 120 kW rack — one rack, the kind our 120 kW AI rack explainer describes — would need hundreds of square metres of radiator, plus the structure, pumps and plumbing to feed it, all launched at thousands of dollars per kilogram. A megawatt-scale facility needs radiator area on the scale of football pitches, and radiators are among the heaviest parts of a spacecraft. That is why Google's Suncatcher prototype runs its TPUs for about 15 minutes at a time: the heat has nowhere to go. Our liquid-cooling explainer sets out how the same heat is removed on Earth — with water and air doing work that, in orbit, only a radiator can do.
Power: sunlight is free, the arrays are the problem
The sun delivers 1,361 watts per square metre at Earth's distance — more energy than humanity uses, many times over. The problem was never the sunlight. It is the hardware between the sunlight and the chip.
A satellite in a typical low orbit circles the Earth roughly every 90 minutes, and spends about a third of each orbit in the planet's shadow. Batteries cover the dark stretch, and batteries are heavy. Google's dawn-dusk orbit avoids most of that, which is why the "up to eight times" figure depends on flying in exactly the right orbit — it is not a property of space in general.
The arrays themselves are the constraint. The GAO's April 2026 assessment says the solar arrays for a large orbital data center would be larger than any arrays previously launched and assembled in space. A NASA study of space-based solar power — beaming power down to Earth, not powering computers in orbit — modelled lifecycle electricity costs 12 to 80 times higher than terrestrial alternatives under its baseline assumptions. None of this says it cannot be done. It says the power system for a large orbital facility is itself a megaproject, and no one has built one.
Getting data up and down
A data center in orbit is useless if it cannot talk to Earth — and talking is harder than it looks.
Latency is not the binding problem. Light travels 500 kilometres in about 1.7 milliseconds, so a signal to a satellite overhead is fast by any human standard. The binding problem is bandwidth: moving the training data up and the results down. AI training data lives on Earth, in data centers and storage systems that are not going anywhere. Uploading it to orbit, and downloading the trained model, means pushing enormous volumes through radio or laser links that are, today, measured in gigabits per second — Axiom's first nodes run 2.5 gigabit-per-second optical links, with 100 gigabits per second planned; Google demonstrated a 1.6 terabit-per-second optical link on a bench, not in orbit. A single terrestrial data-center network moves terabits per second between racks a few metres apart.
That is why the near-term case is data born in orbit: process the imagery where it is collected, and send down the answer. For training a large model on Earth-born data, the round trip is the tax on every byte, and no one has shown a way around it.
What it costs to launch
Everything in orbit starts with a rocket ticket, and the ticket is priced by the kilogram.
A dedicated SpaceX Falcon 9 lists at roughly $70–74 million and can carry up to 22,800 kg to low Earth orbit — which works out to roughly $2,700 to $3,250 per kilogram if you fill it completely. Almost no mission does: that figure is list price divided by maximum payload, a theoretical benchmark rather than a price anyone actually pays. Most small payloads fly rideshare, and SpaceX's published rideshare price for 2026 is $7,000 per kilogram, with a 50 kg minimum. Starship's target of roughly $100–200 per kilogram is a design goal, not a demonstrated price. Google's own paper assumes launch costs could fall below $200 per kilogram by the mid-2030s — an assumption, stated as such, on which much of the economic case rests.
Apply that to a rack. A 120 kW rack weighs tonnes, and it needs solar arrays, radiators, structure and shielding on top. At a few thousand dollars per kilogram, the launch bill alone runs into the millions before the hardware earns a cent — and if anything breaks, no one can drive out to fix it. A failed drive on Earth is a ten-minute swap; a failed component in orbit is a dead satellite, or a retrieval mission that costs more than the hardware.
What would have to be true
For an orbital data center to become more than a demonstrator, five things would each have to happen — and none of them has yet.
1. Radiators that scale. Megawatt-scale heat rejection in orbit, at a mass per kilowatt far below today's ~19 kg per square metre of radiator. Nothing flown so far rejects more than about 100 kW.
2. Launch below a few hundred dollars per kilogram, sustained. Google's paper assumes under $200/kg by the mid-2030s. Today's rideshare price is $7,000/kg.
3. Solar arrays bigger than anything ever assembled in space. The GAO's words, not ours — and assembled, ideally, without astronauts doing it by hand.
4. A workload that belongs in orbit. Data born in space, processed where it is collected. Training on Earth-born data pays the up-and-down bandwidth tax on every byte.
5. Someone to pay for it. No operator has published revenue, a price per GPU-hour, or a customer for orbital AI compute. Starcloud and Axiom describe commercial service from 2027; until a customer pays, the economics are a model, not a market.
Until those five are true, the honest description of the field is the one the GAO gave in April 2026: the technologies exist separately, but deploying and operating them together at data-center scale remains unproven.
Everything in this page is about orbit. For what is actually being built on the ground, see our map of data centers.
Questions readers ask
Isn't space cold, so cooling should be easy?
No — this is the most common misunderstanding in the whole topic, and it gets the physics exactly backwards. Space is a vacuum, and a vacuum is one of the best insulators known: we build thermos flasks out of vacuum to stop heat moving. Without air or water, there is nothing to carry heat away from a hot chip. The only exit is radiation — glowing infrared light into the dark — and at the temperatures electronics can tolerate, a square metre of radiator sheds only a few hundred watts. An orbital data center is a server farm wrapped in a thermos.
Has anyone actually run AI in space?
Yes, in small doses. Starcloud says its single H100 ran a Gemma model and trained a tiny teaching model in orbit in late 2025; Google's prototype returned its first AI response from orbit in October 2026, a haiku from a Gemma model running on a TPU. Both are real flights with real chips doing real (if tiny) work. Neither is a data center: one satellite with one chip, and one prototype running in 15-minute bursts, are demonstrations, not a service.
Why not just put solar panels on Earth instead?
That is the question Google's own team asks, and the answer is the orbit. A panel in a dawn-dusk orbit gets near-constant sunlight and no atmosphere, clouds or night — up to eight times more solar energy per year than the same panel at mid-latitude on Earth, according to Google's paper. On Earth, solar is cheap and proven, which is why terrestrial solar plus batteries is the comparison any orbital proposal has to beat — and, on NASA's 2024 modelling of beaming solar power to Earth, it does not yet beat it.
Will AI training move to space?
Not on any serious scale this decade, on the evidence available. The heat, power, launch-cost and bandwidth problems each have to be solved, and the largest AI training runs need data that lives on Earth. What is plausible sooner is narrower: processing data collected in space, where it is collected, and small constellations serving that work. SpaceX's own IPO filing says orbital data centers "may not achieve commercial viability."
What about the Moon?
The Moon is a different idea from orbit: a surface site, not a satellite. Lonestar's Freedom payload — 8 terabytes of storage — was the first dedicated lunar data-center hardware. It reached the lunar surface in March 2025 aboard Intuitive Machines' IM-2 lander, which tipped over on landing; the payload powered on and achieved partial validation before the power ran out. Lonestar's earlier Independence payload had already tested data storage from the Moon's surface in 2024 aboard Intuitive Machines' IM-1 lander — but as a software test, not a dedicated hardware unit. The Moon's appeal is isolation and cold, dark craters for storage; its problems are the same as orbit's, plus a two-week night, dust, and no one to fix anything.
Sources and further reading
- Google Research: Towards a future space-based, highly scalable AI infrastructure system design (November 4, 2025)
- Google Research: first AI response from orbit (October 30, 2026)
- Google Research: Behind Project Suncatcher (September 24, 2026)
- U.S. Government Accountability Office: GAO-26-109012 (April 28, 2026)
- Thales Alenia Space: ASCEND feasibility study results (June 27, 2024)
- Starcloud company statements via CNBC (December 10, 2025) and TechCrunch (November 2, 2025; A6000 failure)
- Starcloud-1 company account
- Axiom Space: Orbital Data Centers
- Axiom Space: orbital compute update
- Xinhua, CGTN and Sichuan provincial government: Three-Body Computing Constellation (May 2025)
- Data Center Dynamics and IM-2 mission record: Lonestar Freedom lunar payload
- Lonestar: Independence payload
- SpaceX FCC application, DA-26-113 (January 30, 2026)
- SpaceX S-1 (May 20, 2026)
- SpaceX published pricing (2026)
- NASA Ames Research Center: launch-cost estimate (2018)
- NASA: ISS Active Thermal Control System overview (70–100 kW via 422 m²)
- NASA: space-based solar-power study (2024); the original study was not read, figure reported secondarily
