There is something almost comical in seeing the artificial intelligence industry come to grips with an age-old problem: it needs electricity. A lot. We need water, we need cables, land, electrical substations, permits, cooling systems, local communities willing to have huge buildings full of servers working day and night next to them. AI seems intangible when it responds from a screen, yet behind that response is a physical, loud, hot, hungry machine. A car that now someone would like to move outside the Earth.
The idea of data center in space comes from here: taking a part of the infrastructure that today consumes energy and water on the planet and putting it into orbit, where the Sun arrives with more continuity, without clouds, without night in the same way we understand it on earth, and where cooling could exploit the cosmic vacuum. On paper it looks clean, almost elegant. Servers powered by solar panels, optical links between satellites, distributed computing above our heads. Then you look better and the promise becomes rougher: rockets, debris, radiation, costs, impossible maintenance, still vague rules and a quantity of material to launch that makes you thirsty just reading it.
The push comes from data center consumption
According to the most recent estimates from the International Energy Agency, the global electricity consumption of data centers could almost double, going from 485 TWh in 2025 to 950 TWh in 2030approximately 3% of global electricity demand. Data centers dedicated to AI are growing even faster: their consumption could triple in the same period. Inside these numbers there is the less glossy part of artificial intelligence, the one made up of racks, transformers, heat and electrical density. A single advanced rack, about the size of a refrigerator, could reach a peak demand equivalent to that of 65 homes in 2027 and have to dispose of heat equal to that produced by 30 gas boilers.
This is where space begins to seem, at least to some, a shortcut. In orbit the Sun can power panels with greater continuity than on the Earth’s surface. Space data centers would reduce the need for land, fresh water for cooling, and new local power lines. Some companies imagine constellations of satellites capable of working as enormous distributed computing centers, linked together by lasers and powered by solar energy. Google has already presented Project Suncatcheran experimental project to bring satellites with TPU chips and optical links into orbit; the next declared step is the launch of two prototypes by early 2027, in collaboration with Planet.
The race, however, has already taken a form much greater than simple experimentation. In the United States, the FCC, the federal authority that regulates communications and frequencies, has accepted for examination a request from SpaceX for a up to one million satellites intended to function as an “Orbital Data Center”. The proposal talks about orbits between 500 and 2,000 kilometers altitude, high-capacity optical links between satellites and possible connections with the Starlink network. The document even makes reference to a type II civilization on the Kardashev scale, a theoretical classification that measures a civilization based on the amount of energy it manages to use. Translated without science fiction: the ambition is enormous, and so is the encumbrance.
In the meantime, there are those who have already started sending hardware into orbit. Starcloud-1, launched in November 2025, carried an Nvidia H100 GPU into space, a chip used in the most advanced terrestrial data centers. In December, the satellite would run a version of Gemma and train a nanoGPT model in orbit, the company said. These are initial tests, of course, but they mark a concrete transition: the data center in space has stopped being just a conference slide.
Alongside the servers in orbit there is also another idea: using space to power the data centers left on the ground. Meta has announced a partnership with Overview Energy to bring Earth up to 1 GW of space solar powerexploiting satellites in geosynchronous orbit, approximately 35,400 kilometers above the equator, capable of collecting solar energy and sending it to terrestrial photovoltaic systems in the form of near-infrared light. The orbital demonstration is scheduled for 2028, with possible commercial delivery from 2030. Here too the promise is seductive: making existing solar plants produce energy even when the Sun is absent on the ground.
The most intuitive part is energy. Above the atmosphere, solar radiation can be much more favorable than on the ground. A sun-synchronous orbit at the terminator, one that roughly follows the line between day and night, would allow a satellite to keep one side always exposed to the Sun and the other cooler, useful for dissipating heat. So far, everything is pretty tidy. Then comes the dimensions. One scenario cited by the European Parliament Research Service speaks of solar panels up to 4 kilometers per side. For a smaller satellite, equivalent to a rack, much smaller surfaces could be sufficient, around 60 square meters and 28 kW, as happens on systems of the International Space Station. Scale changes everything.
Cooling down is the least romantic part. Saying that space is cold is of little use, because in a vacuum heat is dispersed almost only by radiation. We need circuits with cooling fluids and large radiators aimed at deep space. Those radiators can be comparable in size to solar panels and weigh quite a bit. A leak due to micrometeorites or debris could damage the cooling system and therefore the electronics. Added to this is radiation, capable of altering bits, degrading components and reducing the reliability of chips. We need shielding, redundancy, error correction, fault tolerance. All things that cost weight in space, and weight in orbit costs money.
Maintenance also changes meaning. On the ground, a technician enters the room, replaces a component, checks a cable, works on the system. In orbit, every failure becomes a much drier problem. Intensive AI loads can increase the failure rate of chips; if radiation and thermal cycling are added, the useful life can be shortened. The options remain few: redundancy, replacement of satellites at the end of their life or robotic maintenance, which has yet to be made routine. Communication between satellites itself requires extreme precision: to have high-speed optical links, satellites may have to travel at very close distances, even a few hundred meters. In an already crowded orbit, this is a sentence that should be written slowly.
Then there is the problem of mass. A 1 GW orbital data center, similar in scale to the largest terrestrial facilities under construction, could take longer 10,000 tons of payload to take into space: more than three times the total mass launched into orbit in 2025. A complete data center could require over one hundred launches, followed by annual launches to replace end-of-life satellites. In 2025, all the world’s space launches put together numbered around 300. The idea of moving the problem out of the atmosphere thus begins to bring everything very close to the earth: ramps, fuels, liquid oxygen, emissions, industrial supply chains, port and space infrastructures.
The rules also remain on the ground
The legal issue is less spectacular than a rocket, but it could have the same weight. The Outer Space Treaty, signed in the 1960s, establishes that space does not belong to any state, while launching states remain responsible for space activities. That treaty, however, was born in a world without cloud, without generative AI, without orbiting servers full of personal, industrial, military or health data. When European data is processed in a satellite registered elsewhere, within a constellation managed by a private company, connected to ground stations in multiple countries, the concept of borders becomes less comfortable than usual.
In fact, the idea of a kind of “digital flag” is beginning to appear in the debate, i.e. a legal criterion to establish which law applies to data processed in space. The European GDPR, for example, also regulates data transfers to third countries. A satellite, however, is not a third country in the traditional sense. It is a space object, registered, launched, controlled, connected. Yet it can process enormous amounts of terrestrial data. The gray area is clear and also concerns cybersecurity, liability in the event of an accident, government access to data, relationships between space law and digital law.
On an environmental level, the promise of data centers in space must be handled without sugar. The absence of direct consumption of fresh water in orbit can be an advantage. Abundant solar energy can be even more so. But launches have an impact, satellites at the end of their life re-enter the atmosphere, release materials and can contribute to a form of pollution of the upper atmosphere that is still little understood. Very large constellations increase the risk of orbital congestion, collisions, debris, and interference with astronomical observation. The regulation of low orbits still passes indirectly through the assignment of radio spectrum by the International Telecommunications Union, often according to “first come first serve” logic. For infrastructures of this scale, it seems like little.
The more sober version of the story, then, is this: i data center in space they may become technically possible sooner than it seemed a few years ago. Their main obstacle remains economic, with future costs very much tied to the price of launches and the ability to build, replace and operate orbital hardware on an industrial scale. Some optimistic estimates imagine costs approximately three times higher than those of land-based plants in the short term, with large margins of uncertainty. The fact that something is possible, however, does not automatically make it sensible, nor light, nor clean.
Maybe some of the AI’s computation will actually end up in orbit. Perhaps it will be used first for space missions, Earth observation, data processing near satellites, then for larger loads. Maybe it will remain an expensive niche, good for ads and prototypes. Meanwhile, on Earth, data centers continue to ask for energy, water, space and consensus. Looking up can help you invent something. Escaping from physics, however, is much worse. Even with a rocket.
