Space-Based Data Centers: The Future of Computing Beyond Earth
Orbital data centres promise near-continuous solar power, but vacuum makes cooling hard, radiation damages chips and repair is impractical. We run the numbers on what is real and what is hype.
Data centres are now a visible line item in national energy planning. The International Energy Agency's 2025 analysis of energy and artificial intelligence estimated that global data-centre electricity use could roughly double by 2030, with AI workloads a leading driver. Even single large training runs are substantial: Patterson and colleagues estimated that training GPT-3 consumed on the order of 1,300 megawatt-hours, about the annual electricity use of a hundred or so households, and today's frontier runs are larger. Against that background, the proposal to move part of this computing into orbit, where sunlight is abundant and there is no land to buy, is attracting serious engineering attention. It is also frequently described in ways that skip the hardest physics. This article looks at the numbers.
The attraction: near-continuous solar power
Above the atmosphere, sunlight arrives at about 1,360 watts per square metre, with no clouds, no seasons worth speaking of and, in a dawn-dusk sun-synchronous orbit, almost no eclipse. A triple-junction space solar cell converts roughly 30 percent of that to electricity. A back-of-envelope calculation, my own and purely illustrative, shows that a gigawatt of electrical output would need a few square kilometres of array. That is large but not outlandish compared with the land used by terrestrial solar farms, and the array would produce close to full power around the clock, with no batteries. This continuous generation is the strongest part of the case: a terrestrial solar plant has a capacity factor well below that.
Cooling is not free: vacuum is an insulator
The popular claim that space provides natural cooling is the weakest part of the case. A server on Earth sheds heat by conduction and convection into air or water. In vacuum there is no medium for either, so every watt of electrical power consumed by the processors must ultimately leave as thermal radiation from a radiator. The Stefan-Boltzmann law sets the limit: an ideal surface at 300 K radiates about 460 watts per square metre per side. Real radiators are less than ideal, must keep chips at a modest operating temperature, and in low Earth orbit also absorb sunlight and infrared light from the Earth.
For a facility that converts a gigawatt of electricity almost entirely into heat, this implies radiator area of order a square kilometre or more, plus the pumped-fluid loops needed to carry heat from dense chip packages to those panels. Spacecraft engineers know this problem well; it is why large structures such as the International Space Station carry sizeable radiators. The point is not that cooling is impossible but that it adds mass and complexity, and that the economics depend on mass per kilowatt, not on "free" cold.
Radiation and the hardware problem
Terrestrial processors are protected by the atmosphere and the magnetosphere. In orbit, energetic protons and heavy ions cause two classes of damage: cumulative degradation (total ionising dose) and single-event effects, in which one particle flips a bit or triggers a latch-up. The environment varies with altitude and inclination, and orbits that pass through the South Atlantic Anomaly are more exposed. Barth, Dyer and Stassinopoulos (2003) remains a standard description of this environment.
Radiation-hardened processors exist but lag commercial chips by generations in performance. The alternative is to fly commercial hardware and compensate with error-correcting memory, redundancy and software checkpointing. Hewlett Packard Enterprise's Spaceborne Computer experiments on the International Space Station took this approach with off-the-shelf servers, and showed it can work in a sheltered low-orbit setting. Whether the same strategy holds for thousands of high-power accelerators operating for years in a harsher orbit is exactly what has not yet been demonstrated.
Launch, assembly and the impossibility of repair
Launch costs have fallen with reusable rockets, and the most economical current vehicles are quoted in the low thousands of dollars per kilogram to low Earth orbit. That is still far above the cost of shipping a rack to a warehouse. Every kilogram of solar array, radiator, shielding and structure is a launch bill. A terrestrial data centre swaps a failed server in an hour; an orbital one either carries spare capacity, or has robotic servicing that does not yet exist at this scale, or accepts that failed units stay failed. Orbital debris is a further risk, and a large, low-orbit structure is a large target as well as a large source of future debris if it fails.
Communications and latency, put in proportion
The common argument that latency rules out orbital computing is overstated. Light takes only a few milliseconds to travel between the ground and a satellite at a few hundred kilometres altitude, which is comparable to long terrestrial fibre routes. Geostationary orbit, at about 36,000 km, would add something like 120 milliseconds each way, and that would be a problem. The bigger constraint is bandwidth and the ground link. Training large models needs very high-bandwidth, low-latency links between accelerators; spreading them across many satellites connected by optical inter-satellite links, such as those already flown in commercial constellations, is a different engineering challenge from placing them in one hall. Workloads that move little data but need a lot of computation, or that process data gathered in space, fit the model better.
Where the idea stands
Satellite computing already exists at small scale, in the form of onboard processing for Earth observation and the ISS experiments mentioned above. At larger scale, the idea is at the study and prototype stage. European feasibility work, including the ASCEND study led by Thales Alenia Space, examined whether orbital data centres could be viable, and its reported conclusions were conditional on large reductions in launch emissions and cost. I have not seen peer-reviewed evidence that an orbital facility would beat a terrestrial one on cost per delivered unit of computation today, and claims to that effect should be read as projections.
Limitations and open questions
- Total cost of ownership. No public, independently audited model yet compares mass-launched orbital capacity against terrestrial capacity powered by new nuclear, solar-plus-storage or siting in cold regions.
- Environmental accounting. Rocket exhaust deposits soot and other combustion products in the upper atmosphere, and deorbiting hardware leaves residues. Whether the avoided terrestrial emissions exceed these costs is unresolved.
- Lifetime. Accelerators are replaced every few years on the ground. If orbital hardware must stay in service for a decade to pay back launch costs, it will often be obsolete before it is amortised.
- Governance. Data sovereignty, spectrum, and liability for debris are legal questions that have not been tested for computing infrastructure in orbit.
Questions worth testing
For engineers and analysts, the useful research agenda is quantitative. What is the end-to-end mass per kilowatt of compute, including array, radiator and shielding, and how does it scale? What is the measured upset rate of current-generation accelerators in a representative orbit over a year? What launch price per kilogram makes the break-even point? Can autonomous servicing recover failed modules? Until those figures exist in the open literature, orbital data centres are best treated as a hypothesis with a clear physical basis and an unresolved economic one, not an imminent product.
References and further reading
- International Energy Agency (2025). Energy and AI.
- Patterson, D., et al. (2021). Carbon emissions and large neural network training. arXiv preprint.
- Barth, J. L., Dyer, C. S., and Stassinopoulos, E. G. (2003). Space, atmospheric, and terrestrial radiation environments. IEEE Transactions on Nuclear Science.
- Gilmore, D. G. (ed.) (2002). Spacecraft Thermal Control Handbook. The Aerospace Press.
- Wertz, J. R., and Larson, W. J. (eds.) Space Mission Analysis and Design. Microcosm Press.