Why do tech companies want to build data centers in orbit?

What actually is a data center in orbit?

A data center is, at first glance, a building full of servers that compute, store and deliver. Look at it as an entity made of countless smaller entities — chips, power supplies, cooling loops — that only works because relations to the outside are constantly active: to the power grid, to water or air for cooling, to fiber lines, to land, neighbours and permits. A data center in orbit is the attempt to move that same entity to a place where some of these connections are no longer needed at all.

Concretely, this means satellites or clusters of satellites in low Earth orbit that carry compute chips, draw their power from large solar arrays and send their results back to Earth by radio or laser. The first test satellites with capable AI chips have already launched, and much larger constellations have been announced. A building with foundations becomes a network of flying nodes — the same task, but a completely different web of dependencies.

Why is this being seriously planned right now?

Two developments are converging. First, the energy appetite of AI data centers is growing faster than power grids can be expanded — in many places, the relation between data center and grid has become the most contested connection in the whole system. Wherever a new center wants to plug in, it competes with households, industry and climate targets for the same line. Second, launch costs per kilogram have fallen sharply thanks to reusable rockets, and they keep falling. The connection into space, unaffordable for decades, is just now becoming cheap enough to be worth calculating.

In the model's way of thinking, a balance is shifting. As long as the ‘rocket’ relation was extremely expensive, orbit as a location stayed passive — it existed, but had no active connection to the data center question. The weaker the Earth relations become (scarce power, scarce water, scarce land) and the stronger the space relations get (cheaper launches, better links), the closer the network comes to a tipping point. That is exactly where the plans stand today.

Which connections does orbit cut — energy and cooling?

On Earth, a data center's energy relation is long and indirect: power plant, grid, substations, contracts — many intermediate entities, each with its own bottlenecks. In the right orbit, that same relation becomes short and direct. The sun shines there almost continuously, with no night, no clouds, no atmosphere in between; solar panels deliver several times what they manage on the ground. A contested, shared connection becomes an exclusive one.

Cooling is similar. On the ground it needs water or air conditioning — further active relations to scarce resources. In vacuum there is no air to carry heat away, so what remains is radiation: large radiator surfaces emit the heat into the cold of space. That is technically demanding, but it replaces a relation to a scarce earthly resource with a relation to something nobody misses. This exchange is the core of the whole idea: not more performance, but cheaper connections.

Which new dependencies appear in space?

Cut connections don't simply vanish — new ones take their place, and they are demanding. Every gram of hardware has to ride a rocket; the launch relation replaces the foundation. Cosmic radiation hits the chips without the shelter of the atmosphere, ageing electronics faster and causing compute errors. And nobody can repair anything up there: if a server fails on the ground, a person swaps it in minutes; in orbit the defect stays — the maintenance relation, naturally active on the ground, simply does not exist.

Then there is space debris. Every additional satellite is not only a potential collision victim but future debris itself — a feedback loop in the network: the fuller the orbit, the riskier it becomes for everyone, including your own hardware. And finally, everything hangs on the connection to Earth itself, on radio and laser links to ground stations. Whoever wants to use the compute needs that one relation — it is the bottleneck every result has to pass through.

How do data and results get back to Earth — and what about latency?

Distance itself is less of a problem than many assume. A low orbit sits a few hundred kilometers up — a fraction of a second for a light signal. What matters more is how thick the pipe is: radio and laser links through the atmosphere carry far less than a bundle of fiber in the ground. That is why the first plans do not target everything a data center does, but tasks where a lot is computed and little is transmitted.

Training large AI models is the favourite candidate: tens of thousands of chips compute with each other for weeks, and what comes out at the end is a comparatively compact model. The heavy, permanently active relations then run between the chips in orbit — only an occasional result has to travel to Earth. For anything that needs instant answers for millions of users, ground data centers keep the advantage, because their relations to users are shorter and wider. Zoom out and you don't get a replacement but a division of labour between two network levels.

What does it mean in the larger network — for Earth, energy and society?

Zoom out, and the single data center becomes an entity in a much larger web. On Earth, data centers already compete with cities for power, with agriculture for water and with neighbours for acceptance — all relations that are getting socially tenser. The orbit idea promises to take some of these tensions literally out of the world by outsourcing the hungriest connections to the sun.

Honestly viewed, new tensions appear elsewhere in return. Rocket launches carry their own ecological price, orbits are a shared, finite space, and the question of who owns that space and who may fill it is unresolved. The model offers no verdict here, but a sober checklist: which relations are truly cut, which merely relocated, which newly created — and who carries them? Whether orbital data centers are a relief or just a shift is decided by exactly this balance.

Relations graph

Seen through the model

Imagine training a large AI model. On the ground, that means a building full of chips, a connection to the public grid, cooling towers evaporating water — and a region debating electricity prices. In the model's language: the entity ‘training’ keeps expensive, contested relations active for months, sharing them with everyone else.

Now move the same task into orbit. The chips compute on power that comes straight from the sun and hand their heat to the void. The heavy relations now run exclusively — no grid, no water, no neighbours. In return, new connections have become active that didn't exist on the ground: the rocket that carried everything up, the radiation gnawing at the chips, and the one radio link through which the finished model must eventually travel to Earth.

Whether the move pays off is then not a matter of belief but of balance: are the cut Earth relations more expensive than the newly activated space relations? Today that calculation comes out differently depending on the task — which is exactly why the first projects start with training, not with everyday operations. That is one way to read the development soberly: not science fiction, but shifted connections — a lens, not a finished truth.

Frequently asked

Keep thinking

Related terms: Entity, Relation, The three states: empty, active, passive, Network level, Zoom in / zoom out

• How does a rocket work — as a system of interlocking relations?
• How do vector databases work?
• Is superintelligence only a matter of time?