Data centers for AI in space: is it realistic?

Every time you ask a language model to write an email or generate an image, a huge cluster of servers somewhere on Earth draws as much electricity as a small city. Generative AI is growing so fast that powering data centers has gone from an engineering problem to a geopolitical one. And now an idea that recently looked like pure science fiction is appearing on the horizon: what if we moved data centers into space?

A few figures show the scale of the problem. According to the International Energy Agency, by 2026 the world's data centers will consume more than 1000 terawatt-hours of electricity per year, comparable to the energy consumption of Japan. A significant share of that growth comes from infrastructure for training and inference of large language models.

Companies such as Microsoft, Google and Amazon are buying up nuclear power plant capacity, building their own power facilities and negotiating supply deals with nuclear reactor operators. Even that may not be enough. Pressure from environmental groups and regulators is growing too: data centers not only consume electricity but also give off enormous amounts of heat, and cooling them takes millions of liters of water.

In this context, the idea of orbital data centers stops looking absurd. Its proponents make a simple and elegant case. In orbit, solar energy is available almost around the clock: there are no clouds and no night in the usual sense, and solar radiation is about 40 percent more intense than at the Earth's surface. Cooling, which takes up to 40 percent of the energy that ground-based data centers use, works fundamentally differently in space: radiators dump heat into the cold of open space. Finally, an orbital data center takes up no precious land, makes no noise and does not compete with residential areas for water.

Several companies have moved from theory to concrete projects. The European startup Lumen Orbit has raised funding to develop a prototype orbital computing module. The American company Axiom Space, known for its commercial missions to the ISS, is studying the possibility of placing server racks in space station modules.

Even defense agencies are interested: the Pentagon is funding research into distributed computing in orbit to process data from surveillance satellites. The key factor that moved the idea from fantasy to engineering calculation is the radical drop in launch costs. SpaceX has cut the cost of putting a kilogram into low Earth orbit roughly tenfold compared with the Space Shuttle era, and the reusable Starship system promises to cut it several times more.

But a chasm of technical problems separates an attractive concept from working infrastructure. The first and most obvious is signal delay. Even in a low orbit of 500-600 kilometers, latency is tens of milliseconds, which is acceptable for batch processing or model training but critical for real-time applications.

The second problem is radiation. Cosmic rays and charged particles from the solar wind cause so-called single-event upsets in chips, flipping bits in memory. Server hardware would have to be either shielded, which sharply increases mass, or designed from scratch for radiation tolerance.

The third difficulty is maintenance. When a hard drive fails on Earth, a technician replaces it in minutes. In orbit, replacing a component is a separate space operation costing millions of dollars.

Finally, the growing problem of space debris makes any large-scale orbital infrastructure vulnerable: a collision with a fragment one centimeter across can disable an entire module.

There is also an economic paradox. Proponents of orbital data centers stress the environmental benefits, but rocket launches are far from harmless to the atmosphere. Each Falcon 9 launch releases hundreds of tons of carbon dioxide and soot into the upper atmosphere, where the effects on the ozone layer and climate are poorly understood. If building an orbital computing network takes hundreds of launches, the environmental balance may turn out to be less clear-cut than in startup presentations.

Still, it would be a mistake to dismiss the idea as yet another hype cycle. The history of technology shows that the most radical concepts often find their niche, even if not in the form first imagined. Orbital data centers are unlikely to replace ground-based ones in the foreseeable future, but they could become an important complement for specific tasks: training especially large models, processing satellite data, providing computing to remote regions. A realistic horizon for the first pilot projects is the end of this decade. Until then, the AI industry will keep looking for energy on Earth, glancing upward more and more often, toward the place where the sun never sets.