A single AI training run can consume as much electricity as a hundred homes use in a year. The servers generate so much heat they need water cooling systems that rival small municipal plants. Communities from Virginia to Arizona are blocking new data center projects, worried about power grid strain and water depletion.

This is why some of the world's biggest tech companies are now looking up instead of out. Space-based data centers place computing hardware on satellites in orbit, powered by solar panels that collect unfiltered sunlight 24 hours a day. In February 2025, Elon Musk announced plans to explore orbital computing infrastructure through SpaceX and xAI partnerships. Google has Project Suncatcher exploring similar concepts. China and the European Space Agency have launched feasibility studies.

What sounds like science fiction is becoming an engineering question worth serious money and attention.

How Orbital Computing Works

The core concept involves launching satellites equipped with computing hardware and large solar panel arrays into low Earth orbit, roughly 200 to 1,200 miles above the surface. The satellites receive data from ground stations via laser or radio links, process computations onboard, and transmit results back down.

The appeal is environmental and logistical. Solar panels in orbit collect energy without atmospheric filtering, cloud cover, or nighttime interruptions. A satellite above Earth receives about 35 percent more solar energy than a ground-based panel at peak efficiency. There's no need for cooling towers or water systems because the vacuum of space provides natural thermal radiation. Heat dissipates directly into the void through radiator panels, the same way the International Space Station manages temperature.

Proponents argue this solves three problems simultaneously. It eliminates local resource conflicts by removing data centers from communities. It taps an abundant energy source. It sidesteps the real estate and permitting challenges that slow terrestrial expansion. The infrastructure moves off the grid entirely.

Why Data Centers Became an Infrastructure Crisis

Modern AI models require thousands of specialized processors running simultaneously for weeks or months. Training GPT-4 class models consumes an estimated 10 to 25 megawatts of power. That's enough to run a small factory, operating around the clock. Multiply that by dozens of training runs happening concurrently across the industry, plus the energy needed for inference, and the numbers become staggering.

The cooling problem compounds the power problem. High-performance chips generate intense heat in tiny spaces. Data centers use refrigeration systems, evaporative cooling towers, or direct liquid cooling that circulates water through server racks. In arid regions like Phoenix, a single large data center can consume millions of gallons of water per day, equivalent to thousands of households.

Local opposition has grown fierce. Residents near proposed data center sites in Loudoun County, Virginia have organized against projects they view as infrastructure leeches. The facilities bring few jobs compared to their resource demands. They strain electrical grids during peak hours. The NIMBY dynamic is slowing approvals and driving up costs for tech companies that need to expand capacity quickly.

The Technical Barriers No One Has Solved

Latency is the killer constraint. Light travels fast, but not instantaneously. A signal traveling from a ground station to a satellite in low Earth orbit and back covers roughly 500 to 2,500 miles, depending on satellite position and ground station location. That round trip introduces 5 to 30 milliseconds of latency under ideal conditions, not accounting for atmospheric interference or signal processing delays.

For many AI workloads, this matters enormously. Training large language models requires millions of rapid-fire data exchanges between processors. Each training step depends on the results of the previous one. Adding even 10 milliseconds per exchange would slow training runs from weeks to months or make them computationally impractical. Real-time inference tasks like chatbots, voice assistants, and autonomous vehicle decision making cannot tolerate multi-second delays while data travels to orbit and back.

The Cooling Paradox

Thermal management in orbit is more complex than it appears. Yes, space is cold, but there's no air to carry heat away. Spacecraft rely entirely on thermal radiation, which means large radiator panels to dissipate heat from densely packed processors. The International Space Station's radiators span 1,700 square feet to cool a fraction of the power load a modern data center generates. Scaling this to handle megawatts of computing heat requires radiator systems so large they might dwarf the computing hardware itself.

Satellites in low Earth orbit move at roughly 17,000 miles per hour, circling the planet every 90 minutes. Maintaining constant communication with ground stations requires either constellations of dozens of satellites so one is always overhead, or higher orbits where satellites move slower but latency increases. Launch costs, while declining, still run thousands of dollars per kilogram of payload. A single rack of high-performance servers weighs over 1,000 pounds before adding solar panels, radiators, and communications equipment.

What This Is Actually Good For

Space-based computing isn't viable for latency-sensitive workloads, but it could excel at specific batch processing tasks that don't require real-time responses. Training certain types of models where data can be uploaded in bulk, processed over hours or days, and results downloaded later might work. Climate modeling projects like the European Centre for Medium-Range Weather Forecasts' seasonal prediction models, which process historical atmospheric data over multiple days to generate three-month forecasts, could tolerate orbital delays. The same applies to protein folding simulations and large-scale data analysis jobs that already run overnight on terrestrial servers.

Archival storage represents another potential use case. Data that needs to be retained but rarely accessed could sit on orbital storage satellites, retrieved only when necessary. The latency penalty matters less when retrieval times measured in minutes are acceptable.

Specialized edge processing for satellite networks might close the loop more elegantly. If data is already being collected by Earth observation satellites, processing it in orbit before sending condensed results to the ground reduces bandwidth needs. This works for applications like agricultural monitoring, disaster response analysis, or environmental tracking where time scales are measured in hours or days, not milliseconds.

The Companies Making Real Moves

Several major technology and aerospace organizations are actively exploring orbital computing infrastructure, though no operational systems have been deployed yet.

SpaceX and xAI have announced plans to explore orbital computing infrastructure, positioning it as central to long-term AI development strategy. The announcements leverage SpaceX's launch capabilities and satellite experience from Starlink, though concrete timelines and technical specifications remain undisclosed.

Google's Project Suncatcher explores solar-powered orbital computing with a focus on sustainability metrics. The project remains in early research phases, with no announced launch dates. Internal presentations reportedly emphasize reducing data center carbon footprints by moving compute-intensive training runs to space-based platforms powered entirely by solar energy.

China's space program has allocated funding for feasibility studies on orbital data infrastructure as part of its broader space station and satellite network ambitions. European Space Agency documents reference edge computing in orbit as a long-term research area, integrated with Earth observation satellite programs already in operation.

The pattern is consistent across all initiatives: ambitious concepts backed by preliminary research, but no operational hardware in orbit or confirmed deployment schedules.

What This Means for Infrastructure Decisions

Moving data centers to orbit doesn't address the underlying question: who gets a say in infrastructure decisions that affect shared resources? The terrestrial data center fights revealed something deeper than NIMBY politics. Residents were asking whether their local power and water should subsidize global AI infrastructure that brings minimal local benefit. Space-based systems remove the immediate environmental burden but also remove any community participation in governance or benefit-sharing.

Labor implications remain murky. Data centers employ thousands of technicians, engineers, and operations staff. Orbital systems would require far fewer hands-on workers, with most operations managed remotely from ground control facilities. The infrastructure jobs shift from distributed local employment to concentrated technical roles at launch sites and control centers.

Progress That Lasts Requires Honest Tradeoffs

Space-based computing addresses real problems. Terrestrial data centers strain resources and face opposition. Solar power in orbit is abundant and constant. The environmental case has merit.

But the technical challenges are not yet solved, the economics remain speculative, and the social implications deserve scrutiny beyond corporate press releases. The question is not whether we can put data centers in orbit. The question is whether we should, for which workloads, and at what cost to whom.

Right now, we have more announcements than answers, more ambition than evidence, and more engineering problems than operational systems. That's not a reason to dismiss the concept. It's a reason to watch the evidence closely and ask hard questions about feasibility, accountability, and who benefits when infrastructure moves beyond democratic reach.

Every watt is a choice. So is every orbit.