The relentless appetite of artificial intelligence for computing power has pushed some of the world’s largest technology companies to consider an idea once confined to science fiction: building data centres not on land, but in orbit. As power grids strain and land grows scarce for sprawling server farms, companies from SpaceX to Google are now racing to put racks of GPUs into low-Earth orbit, betting that the vacuum of space can solve problems the Earth no longer can.
Why Space Suddenly Makes Sense
The logic is straightforward even if the engineering is not. Terrestrial data centres depend on three constrained resources — land, power and water — and AI’s compute demands are outpacing all three. Tech companies are exploring orbital data centers to power AI from space, envisioning networks of thousands of satellites that can complete the computations needed to sustain AI’s growing energy demands.
Global data-centre power consumption is expected to roughly double to nearly 1,000 terawatt-hours by the end of the decade, according to the International Energy Agency. That surge has triggered community backlash worldwide over electricity bills and water use, pushing firms to look skyward.
Elon Musk’s SpaceX has been the most aggressive mover, filing with regulators to deploy up to a million satellites that would operate as orbital data centres. Google, meanwhile, is pursuing its own moonshot. Project Suncatcher envisions an 81-satellite cluster built in partnership with satellite-imagery company Planet, with two prototype satellites due to launch in early 2027. Nvidia-backed startup Starcloud has already gone further, launching a satellite carrying an Nvidia H100 GPU that successfully trained and ran a version of Google’s Gemma AI model in orbit, marking the first AI model training in space.
The appeal is largely about energy economics. Satellites in certain orbits can receive near-constant sunlight, potentially allowing data centres to run on clean solar power without interruptions from night or weather.
India Joins the Space Race for Compute
For India, the stakes are especially pointed. Despite hosting almost 20 percent of the world’s data, India has only 3 percent of global data centre capacity, a gap officials view as a strategic vulnerability as the country pursues its AI ambitions.
Domestic players are responding quickly. Bengaluru-based cloud firm NeevCloud has begun testing India’s own orbital edge infrastructure, with founder Narendra Sen arguing that in space there is unlimited access to solar power and excess heat can simply be radiated away, eliminating the water and energy problems that plague terrestrial facilities. The company’s pilot launch is scheduled before the end of 2026, with an ambitious goal of scaling to more than 600 orbital edge data centres over the following three years.
Separately, satellite-imaging startup Pixxel Space has teamed up with Sarvam AI to build an orbital satellite named Pathfinder. The collaboration aims to process large volumes of space-based data in real time using advanced AI models, with Sarvam’s models analysing crops, infrastructure and weather patterns directly in orbit. Reports also suggest the Department of Space and ISRO are separately examining whether orbital facilities could process satellite and communications data domestically, reducing reliance on foreign infrastructure for sensitive government workloads.
The Hard Physics That Could Ground the Dream
For all the enthusiasm, the technical hurdles remain formidable. Unlike on Earth, satellites cannot shed heat through air convection — they must rely on radiators alone, a far less efficient process in the vacuum of space. Cost is another obstacle: orbital data centres are currently estimated to be about three times more expensive than their terrestrial counterparts once launch, maintenance and replacement expenses are factored in.
Radiation exposure, orbital debris risk, and an almost entirely undefined legal framework around data sovereignty in space add further complications. Industry analysts broadly agree that meaningful, revenue-generating orbital compute is unlikely before the late 2020s, and full-scale deployment may take a decade or more to mature. Whether this becomes the next great infrastructure shift or another expensive detour will depend on how quickly launch costs fall — and how urgently Earth’s own power grids run out of headroom
