Why tech companies want to put data centres in space
AI's hunger for electricity is running into the limits of power grids. Orbit offers constant sunlight, and a long list of engineering problems.
Next week, Google launches a satellite carrying four of its AI chips. The reason is electricity.
The problem: AI runs on power
Training and running large AI models happens in data centres packed with specialised chips. Those chips draw huge amounts of electricity, and they turn nearly all of it into heat, which then takes more electricity to remove.
As companies race to build more AI capacity, the bottleneck is increasingly not chips or buildings but power: finding enough of it, connecting it to the grid and doing so without pushing up emissions. New data centre projects around the world now wait years for grid connections.
The pitch: the sun never sets in the right orbit
On the ground, solar panels deliver power only in daytime, less on cloudy days, and lose some of the sunlight to the atmosphere. In certain orbits, a satellite can sit in near-constant sunlight.
Put those advantages together and a solar panel in space can collect several times more energy over a year than the same panel on the ground. Reporting on Google’s project puts the figure at up to eight times.
If you could put the chips next to the panels, the argument goes, you would have an AI data centre powered by an almost limitless, free energy source.
The catches
Nobody thinks this is easy. The main challenges:
Heat. On Earth, data centres use air and water to carry heat away. In the vacuum of space there is no air, so heat can only leave by radiating away, which requires large radiator panels.
Radiation. Outside the protection of Earth’s atmosphere, chips are hit by charged particles that can flip bits, cause errors and slowly degrade hardware.
Networking. Big AI jobs need thousands of chips talking to each other at high speeds. Doing that between satellites means flying them in tight formation and linking them with lasers.
Getting data up and down. Models and results still have to travel between orbit and the ground.
Repairs. When a server fails on Earth, a technician swaps it out. In orbit, it stays broken.
Launch cost. Every kilogram sent to orbit costs money, though reusable rockets have brought that cost down dramatically.
So is it realistic?
Right now, it’s an experiment. Google’s first satellite is a test of whether its chips survive and work at all. Commercial space data centres at meaningful scale would need progress on almost every item above.
But the underlying pressure, AI’s appetite for power, is real and growing. That’s why serious engineering money is now going into ideas that sounded absurd a few years ago.
