As AI systems grow, so does the question of where to plug them in. A handful of companies now want to put AI data centers in orbit, where sunlight is nearly constant and there is no local grid to wait on, and small test satellites are already flying. Whether data centers in space can ever compete with ones on the ground comes down to two engineering problems: getting enough power, and getting rid of the heat that power turns into.
Why Put AI Data Centers in Space?
On Earth, data centers already compete for electricity, water, and land. The International Energy Agency estimates that cooling uses about 7% of a data center’s electricity at efficient hyperscale sites and more than 30% at less efficient enterprise sites (IEA, Energy and AI), and local grids can become a bottleneck long before chips do — the story behind AI’s thirst for watts and water on Earth. Reasoning models push demand further, because they can spend many times more computing on a single hard answer (see how test-time compute works).
Space changes the math on paper. Google says a solar panel in the right low Earth orbit gets near-constant sunlight and can produce up to eight times more power than the same panel on Earth (Google). There is no grid connection queue, and the cold background of space is where waste heat has to go anyway. The pitch is that abundant solar power could support AI workloads that ground sites struggle to host — if mass, reliability, and dollars per watt cooperate.
That “if” matters. Marketing often pitches “unlimited solar.” Engineering still sizes arrays for peak load plus eclipse reserve, then pays launch and structure costs for every square meter of panel and radiator.
How Much Power Does an Orbital Data Center Need?
Modern training and inference clusters are measured in megawatts. Moving that demand off-planet does not erase it; it relocates it. Solar arrays must cover peak load and night passes. Batteries or other storage must ride through eclipses and faults. Radiation-tolerant power electronics add mass and cost. Every extra kilowatt of sustained draw means more panel area, more structure, and more launch mass — costs that scale faster than many early slide decks admit.
Published designs give a sense of the scale. A 2026 engineering analysis estimated that each megawatt of computing in orbit would need about 5,600 square meters of solar panels at the start of the mission, plus batteries for eclipses and radiators for heat (arXiv:2604.27197). SpaceX’s first data center satellite design, called AI1, would carry about 600 square meters of solar panels producing roughly 150 kilowatts at peak and 120 kilowatts on average, according to figures Ars Technica took from a SpaceX schematic. Starcloud has asked the FCC for permission to operate 88,000 spacecraft (TechCrunch).
That is why panel efficiency matters more in orbit than on a rooftop. Higher-efficiency cells such as perovskite-silicon tandem solar cells promise more watts per square meter, which counts most when every square meter has to be launched.
The Cooling Problem: Getting Rid of Heat in a Vacuum
In a vacuum there is no air or water to carry heat away, so an orbital data center has to send all of its waste heat into space as infrared light from radiator panels. Almost every watt the chips use ends up as heat, so more computing means more radiator area and more mass. That is why cooling, more than chips or radiation, is often called the hardest part of building data centers in space (Ars Technica).
Physics sets the ceiling. Under the Stefan–Boltzmann law, a perfect radiating surface at 350 kelvin (about 77°C) gives off roughly 850 watts per square meter. Real panels do worse, because they are not perfect emitters, they partly see other parts of the spacecraft, and they soak up heat from sunlight and from Earth. The arXiv analysis works out to about 1,500 square meters of radiator per megawatt before that outside heating is counted, and about 2,500 square meters per megawatt in its base case once it is (arXiv:2604.27197). Running radiators hotter shrinks them quickly, because emitted power rises with the fourth power of temperature, but the chips feeding them have to stay cool enough to work. Starcloud says the link between chip temperature and radiator size is one of a few key design choices its engineers are tracking (TechCrunch).
The International Space Station is the best real-world benchmark. Its six ammonia-cooled radiators weigh just over 6 metric tons together and shed about 70 kilowatts of heat, which Ars Technica notes is roughly in line with what SpaceX’s proposed satellites would need (Ars Technica). Starcloud’s chief executive told Ars that two-thirds of the company’s engineering team is working on a cheaper, lighter radiator. Big radiators also need protection, because panels in low orbit get struck by tiny, fast debris, the same threat that thin shields that stop micrometeoroids were designed for.
Who Is Building Data Centers in Space (October 2026)
Real hardware is flying, though at kilowatt scale, not data-center scale. Here is where the main efforts stood as of October 8, 2026.
| Who | What has flown | What’s next | Sources |
|---|---|---|---|
| Starcloud | Starcloud-1, a 60 kg satellite with one Nvidia H100 GPU, launched November 2, 2025 | Two 8 kW Starcloud-2 satellites on 2027 rideshares; the larger Starcloud-3 is meant to fly on Starship | DCD; CNBC; TechCrunch |
| Google Project Suncatcher | A prototype satellite carrying Google TPUs, built with Planet, launched October 1, 2026 | Two satellites in 2027 to test laser links between them | Google; Google |
| Axiom Space | Orbital Data Center Nodes 1 and 2, launched January 11, 2026 with Kepler’s optical relay satellites | Grow the network from kilowatts toward megawatts by 2030 and beyond | Axiom Space |
| SpaceX | Nothing yet | Applied to the FCC on January 30, 2026 for up to 1 million data center satellites; still under review | FCC; SatNews |
On October 1, 2026, SpaceX carried a prototype satellite for Google’s Project Suncatcher into orbit on the Transporter-18 rideshare. The satellite was built in partnership with Planet. It carries Google’s tensor processing units, the chips the company uses for artificial intelligence. Google said its team has confirmed contact and that the satellite is operating as expected. Over the coming weeks the team plans to gather data on how those chips handle the physical stress of spaceflight, plus the radiation and thermal extremes of space. Google calls the flight the first step in a long research effort on whether space could one day hold large machine-learning systems. The point of this flight is to test the chips in orbit, not to open a data center.
Starcloud’s first satellite is the closest thing to a working orbital data center so far. Starcloud-1 is about the size of a small refrigerator, flies roughly 325 kilometers up, and carries an Nvidia H100, a data-center-class GPU (DCD). In December 2025 the company said it had trained a small language model, NanoGPT, in orbit and was running Google’s open Gemma model there (CNBC). In August 2026 it raised $250 million at a $2.3 billion valuation, with Nvidia and Cisco among the investors, and its next step has slipped: TechCrunch reported both the funding and that two 8-kilowatt Starcloud-2 satellites are now planned for rideshare launches in 2027.
Axiom Space launched its first two Orbital Data Center nodes on January 11, 2026, alongside the first batch of Kepler Communications’ optical relay satellites, after testing a prototype unit (AxDCU-1) on the International Space Station in 2025 (Axiom Space). These nodes are built to process data from other spacecraft in orbit, not to replace cloud campuses on the ground.
SpaceX’s plan is the largest by far. On January 30, 2026, it asked the U.S. Federal Communications Commission for permission to launch and operate up to one million satellites, flying between 500 and 2,000 kilometers up, as an “Orbital Data Center System” (FCC public notice DA 26-113). The FCC accepted the application for filing on February 4, 2026, which opens it for public comment but does not approve it. In September 2026, SpaceX was still answering FCC questions about debris and re-entry risk, and the application remained under review (SatNews).
Flight plans and performance claims shift quickly. Treat press milestones as progress markers, not guaranteed megawatt capacity.
Are Space Data Centers Feasible? Launch Cost, Radiation, Latency, Debris
Launch is the first test. Ars Technica’s rough estimate for SpaceX’s million-satellite plan assumed satellites that last about five years and found it would take at least 3,500 launches a year even in the most optimistic case, compared with 329 orbital launch attempts worldwide in 2025, at an all-in cost of roughly $1.45 trillion to $9.8 trillion (Ars Technica). Ars concludes that none of it looks feasible unless Starship cuts launch costs dramatically.
Radiation looks more manageable than once feared. Google says its Trillium TPUs survived a proton-beam test at a higher total radiation dose than they would receive in a five-year space mission (Google), and Starcloud says its H100 has performed well in orbit so far (Ars Technica). Repairs in orbit are unlikely, so these satellites would probably be replaced after about five years, which is also roughly how long a chip stays cutting-edge.
Latency and bandwidth add another tax. Training giant models in orbit still needs data movement and ground coordination. Near-term value is clearer for edge workloads — filtering satellite imagery, running onboard autonomy, or keeping sensitive inference off terrestrial networks — than for replacing a hyperscale Earth campus overnight.
Debris is the shared risk. Up to a million large satellites would add to an orbital environment that is already crowded, raising the stakes of the Kessler syndrome, where collisions create debris that causes more collisions.
Megawatts and Beyond: Solar Arrays vs. Nuclear in Orbit
Longer term, orbital megawatt-class AI would require either vast solar farms with high-voltage distribution or next-generation sources such as compact nuclear power for continuous baseload. Lunar concepts must solve two-week nights and dust. In both cases, economics hinge less on peak chip FLOPS and more on dollars per delivered watt-hour over the mission life — including replacement, radiation degradation, and end-of-life disposal or deorbit.
Nuclear is slow even on the ground: TVA’s small modular reactor construction permit arrived in September 2026, and that reactor is not built yet. A compact reactor in orbit would face that same licensing and engineering path, plus a launch.
Software will matter as much as hardware. Workloads that can throttle with eclipse, prioritize inference over training when power dips, and checkpoint cleanly will survive; rigid always-on training schedules will not. Matching compute to the energy budget is the unglamorous core of “AI in space.”
What to Watch in 2027
These signals will show whether data centers in space are maturing:
- Specific power: demonstrated kilowatt-to-megawatt solar or nuclear systems with high watts per kilogram after radiation and thermal margins.
- Thermal packaging: radiators and cold plates that keep accelerator fleets in limits without ballooning mass.
- Power-aware software: schedulers that throttle inference and training as eclipse, dust, or faults change the energy budget.
- Flight results: data from Google’s Suncatcher prototype and its planned two-satellite laser-link test in 2027 (Google), and from Starcloud’s 8-kilowatt Starcloud-2 satellites, now planned for 2027 (TechCrunch).
- The FCC’s decision: whether regulators approve SpaceX’s million-satellite application, and what debris rules come attached.
Also watch optical intersatellite links and ground-segment cost. Cheap chips in an expensive truck still lose. Cheap launch helps, but it does not delete radiator area or battery cycle life.
AI in space will not escape physics. It will succeed when power architecture, thermal design, and launch economics move together — not when chips alone get faster. For now, the honest story is kilowatt prototypes, not a sudden orbital replacement for Earth hyperscale campuses.
Sources and further reading: Ars Technica: How hard is it to build orbital data centers, actually?; Orbital Data Centers: Spacecraft Constraints and Economic Viability (arXiv); Google: Behind Project Suncatcher; FCC public notice on SpaceX’s application.
Further reading
P3 Kill A Watt Electricity Usage Monitor
A plug-in meter that shows the watts and kilowatt-hours your own devices draw. A small, hands-on way to feel the power math that decides where big compute can live.
ARCTIC MX-4 Thermal Paste (4 g)
The everyday version of the cooling problem: a carbon-based paste that moves heat from a CPU or GPU into its heatsink. Every watt a chip uses ends up as heat that has to go somewhere, on Earth or in orbit.
Etekcity Lasergrip 1080 Infrared Thermometer
Point and click to read surface temperatures from a distance. Handy for spotting hot laptops, chargers and vents, and a hands-on way to see how heat builds up when it can’t escape.
