At a March event, Elon Musk sketched a bold vision: put data centers into orbit. After SpaceX’s reported merger with xAI, Musk argued that space could solve Earth’s power constraints — sunlight is plentiful up there, he said — and suggested swarms of AI-running satellites could eventually be cheaper than terrestrial centers within a few years.
Experts say the idea is intriguing but the timeline is optimistic. Brandon Lucia, a Carnegie Mellon professor who researches satellite computing, calls the basic math appealing, but points to major engineering hurdles. Abundant sunlight doesn’t eliminate the hard problems of getting usable power to processors, shedding heat in vacuum, networking huge volumes of data between moving platforms, and launching vast amounts of hardware into space.
Why consider orbit at all?
AI’s appetite for electricity is exploding. The International Energy Agency projects global data-center energy use could nearly double to around 1,000 terawatt-hours by the end of the decade. Companies are building new power plants and exploring nuclear and gas options to keep up. For some entrepreneurs, orbital data centers are a way to sidestep terrestrial siting and grid limits and tap near-constant solar energy.
Startups and incumbents are already testing pieces of the idea. Philip Johnston, CEO of Starcloud, which aims to build orbital compute platforms, warned that constraints on where new power facilities can be sited on Earth could leave chips idle. Starcloud launched an early demonstrator carrying an Nvidia H100 and ran a version of Google’s Gemini from space; a follow-up craft planned for October will produce about 8 kilowatts — far below data-center scale but useful as a technical demo.
Google is exploring the concept too. Project Suncatcher imagines an 81-satellite cluster built with Planet, with two prototypes planned for early 2027. Planet’s CEO says orbital data centers are “an idea whose time has come,” though the precise economics remain unsettled.
Scale, power and the ISS yardstick
To understand the scale, compare to the International Space Station. The ISS generates roughly 100 kilowatts from solar panels that cover about half a football field — comparable to the output of a large car engine. An average large terrestrial data center might draw tens to hundreds of megawatts. Building a 100-megawatt facility in orbit, by some estimates, would require solar arrays and structures hundreds to a thousand times larger than the ISS footprint, depending on orbital choices. That’s physically possible in principle, but assembly, launch and integration on the scale required are enormous undertakings that don’t mesh easily with a two- or three-year timeline.
Cooling in a vacuum
Space isn’t inherently cold for electronics. Because there’s no air to carry heat away, processors can overheat unless systems radiate heat into space. Satellites use pumped fluids and large radiator panels to reject thermal energy; an orbital AI node would need proportionally massive radiators in addition to solar arrays. That pushes designers toward either very large single platforms or huge constellations of smaller craft with distributed radiators.
Constellation trade-offs
Smaller satellites flown tightly together could spread power generation and thermal loads across many vehicles, but they create new demands: very high-bandwidth inter-satellite links, likely lasers, to move data between nodes. Even with light-speed links, distances and routing can add latency that hampers certain AI workloads. Google’s concept proposes tightly clustered satellites to minimize that latency; other plans call for dense global constellations, including polar coverage to reach all regions.
Launch costs and logistics
Current launch prices are roughly $1,000 per kilogram to low Earth orbit. Google’s internal estimates suggest costs need to fall to about $200/kg to make space data centers economically attractive. Musk hopes SpaceX’s Starship — a heavy-lift reusable rocket still under development — will sharply reduce launch costs if it succeeds, but Starship’s timelines and performance remain uncertain. Even with lower per-kilogram prices, putting up the vast arrays and radiators needed will be expensive.
Operational realities and maintenance
Terrestrial data centers are highly serviceable: technicians, vendors and customers routinely visit facilities to install, upgrade and replace hardware. Raul Martynek, CEO of DataBank, notes that his Ashburn, Virginia, site consumes about 13 megawatts continuously — roughly 130 times the ISS — and relies on hands-on operations. Space platforms could lean heavily on software robustness, exhaustive prelaunch testing, and future robotics for servicing, but many enterprise customers demand physical control, rapid replacements and on-site audits.
Who could win, and when?
If launch costs fall, large radiators and solar arrays become practical, inter-satellite networking matures, and on-orbit servicing becomes routine, orbital data centers could open a new frontier in computing. Companies like Starcloud, Google’s Project Suncatcher, Planet and several startups are building prototypes and proving elements of the concept — running limited AI models from low-power satellites and demonstrating high-speed links.
For now, most experts treat orbital data centers as a long-term possibility rather than an imminent disruption. The physics of power generation, heat rejection and latency, combined with launch economics and maintenance realities, make space-based AI infrastructure technically challenging and capital-intensive. ‘‘Not next year and certainly not in three years,’’ MIT’s Olivier de Weck says — a caution echoed across the field.
Contact Geoff Brumfiel on Signal: gbrumfiel.13