A typical AI-focused data centre can consume as much electricity as 100,000 homes. Now companies want to put them in orbit, where carefully chosen satellites can see almost continuous sunlight and dump their waste heat into space through giant radiators instead of conventional cooling plants.

A typical AI-focused data centre now belongs in the same power conversation as heavy industry. The International Energy Agency uses a 100-megawatt hyperscale AI facility as its comparison point, equivalent to the annual electricity consumption of roughly 100,000 households. It expects worldwide data-centre electricity demand to more than double from about 415 terawatt-hours in 2024 to around 945 terawatt-hours in 2030.

That growth is colliding with grid queues, transformer shortages, local opposition and the water demands of some cooling systems. The proposed escape route sounds like science fiction: put part of the computing industry in orbit. In a carefully chosen orbit, solar panels can see the Sun almost continuously and computers can reject heat through radiators instead of cooling towers and chillers.

The physics allows it. The difficult question is whether an entire data centre, including its power, thermal control, communications, shielding and replacement plan, can close technically and economically at useful scale.

The attraction begins with near-continuous sunlight

A dawn-dusk sun-synchronous orbit keeps a spacecraft close to the moving boundary between Earth’s day and night. The orbital plane turns gradually as Earth goes around the Sun, allowing suitably oriented solar arrays to remain illuminated for most of nearly every orbit. “Near-continuous” matters: eclipses can still occur depending on altitude, geometry and season, and batteries or reduced workloads may still be needed.

Google’s Project Suncatcher proposes compact constellations carrying Tensor Processing Units and connected by optical links. Google says a panel in the right orbit can be up to eight times more productive than one on Earth because it avoids night, clouds and atmospheric losses. Its illustrative analysis models 81 satellites flying as a cluster at about 650 kilometres altitude. A two-satellite learning mission with Planet is planned for early 2027.

Other efforts have moved beyond slides. Starcloud-1 launched in November 2025 carrying an Nvidia H100 GPU, then ran a version of Gemini and trained a small language model in orbit. In March 2026, Nvidia announced a space-computing product line and named Aetherflux, Axiom Space, Kepler Communications, Planet, Sophia Space and Starcloud among companies using its hardware for orbital or space-edge computing.

Space is a heat sink only through radiation

The phrase “cold of space” hides the central engineering problem. A vacuum contains almost no matter, so it cannot carry heat away by convection. Nearly every watt fed to a processor ends up as heat. Engineers must conduct that heat away from the chips, spread it through fluid loops or heat pipes, and emit it as infrared radiation from panels with a clear view of deep space.

Radiators work, but their output depends steeply on temperature. Hotter radiators can reject more heat per square metre, yet processors, memory, pumps and materials impose temperature limits. Solar arrays, warm Earth and direct sunlight can also add heat, so radiator orientation and surface coatings matter.

A 2026 preprint, Orbital Data Centers: Spacecraft Constraints and Economic Viability, modelled a representative one-megawatt high-sunlight system. Slava G. Turyshev calculated about 5,640 square metres of beginning-of-life photovoltaic area and roughly 2,500 square metres of radiator area, before accounting for all fixed spacecraft hardware. A simple linear scale toward 100 megawatts would imply hundreds of thousands of square metres of deployed surface. An actual design would not be a simple scale-up, but the comparison shows why radiators are infrastructure, not accessories.

Orbit could avoid cooling towers that consume fresh water and could remove conventional chillers from the design. It does not provide free cooling. Every radiator must be manufactured, folded into a rocket, deployed without failure and protected against thermal cycling, micrometeoroids and debris.

A data centre is also a communications machine

Power and cooling are only two parts of the system. Large AI workloads divide calculations among thousands of accelerators, which must exchange intermediate results quickly and predictably. Google’s analysis says matching terrestrial data-centre networking would require inter-satellite links carrying tens of terabits per second. Its concept therefore places satellites only hundreds of metres apart and uses free-space optical connections.

That tight formation creates its own demands. The nodes must point lasers precisely, maintain relative position and coordinate work despite vibration, drift and hardware faults. A cluster can be modular, but software must recover when a satellite fails or moves out of formation.

The path to users on Earth is harder still. Optical downlinks can offer high capacity, but clouds interrupt them. Radio provides greater weather tolerance at lower bandwidth. A general cloud service may need networks of ground stations, storage to buffer interrupted links and hybrid optical-radio communications. Latency from low Earth orbit can be modest, but capacity and availability are more important than propagation time alone.

The first useful workloads may already begin in space

The strongest early case is not moving every web search or chatbot query off Earth. It is processing data where that data is collected. Imaging and radar satellites produce large raw streams but users often need only a fire alert, a storm track, a crop classification or a compressed scientific product. An orbital computer can turn raw observations into a much smaller result before transmission.

That logic also underpins earlier European Space Agency studies of space-based data centres. ESA examined architectures in which an observing satellite passed data to another spacecraft for processing, reducing the amount that had to cross the space-to-ground bottleneck. Such space-native edge computing can create value with kilowatts or megawatts, well before anyone builds an orbital equivalent of a 100-megawatt campus.

Model inference may fit before model training for the same reason. Training repeatedly moves huge datasets and parameters between processors. Inference can use a model already stored aboard the spacecraft, accept a relatively small input and return a compact answer.

Launch cost must fall while reliability rises

Every photovoltaic blanket, radiator, shield, optical terminal and processor must survive launch. Google’s economic analysis suggested that if launch prices to low Earth orbit fell below about $200 per kilogram by the mid-2030s, launching and operating orbital compute could approach reported terrestrial data-centre energy costs on a per-kilowatt-year basis. That is a conditional projection, not a current price.

Turyshev’s independent model reached a more demanding boundary. At around 40 kilograms of deployed system per delivered kilowatt, a terrestrial infrastructure benchmark of $10,000 to $40,000 per kilowatt would permit only about $250 to $1,000 per kilogram for spacecraft construction and launch combined. Communications, operations, utilisation and finite service life would tighten the constraint further.

Space hardware must also tolerate ionising radiation and single-event errors. Google exposed its Trillium TPU to a proton beam and reported no hard failures attributable to cumulative ionising dose through the maximum test, although high-bandwidth memory showed the earliest irregularities. Shielding and error correction help, but shielding adds mass and does not eliminate every upset.

Orbit exchanges terrestrial constraints for orbital ones

The United States Government Accountability Office reported in 2026 that public and private projects were testing relevant hardware and that the Federal Communications Commission had received three applications for large US data-centre constellations since January. It also highlighted collision risk, astronomical interference, cybersecurity, uncertain environmental effects and economics as unresolved issues.

Large arrays and radiators increase the cross-sectional area exposed to debris and residual atmospheric drag. They can reflect sunlight into telescopes, require collision-avoidance manoeuvres and eventually need controlled disposal. Repairs that are routine in a ground facility may require robotic servicing or replacement of an entire spacecraft.

The proposal therefore deserves neither dismissal nor the claim that space makes energy and cooling free. As an earlier SpaceDaily analysis of orbital compute economics explained, solar flux alone does not close the system. The likely path begins with small, specialised processors beside space-based sensors, followed by demonstrations that prove reliable networking and thermal control. Only then will the industry know whether giant orbital data centres are a practical extension of the cloud or an elegant idea defeated by the mass of everything surrounding the chips.