A Satellite Runs AI in Orbit and Identifies What It Sees Before Sending Anything Down

Planet's Pelican-4 spacecraft spotted and labeled objects in its own images while still in space, using a small Nvidia processor onboard. It cuts the wait for an answer from hours to minutes.

A satellite operated by Planet, the company that photographs the entire surface of the Earth every day, successfully ran an AI program onboard for the first time this week. The spacecraft, called Pelican-4, examined its own images while still in orbit, identified objects in them with about 80 percent accuracy on unprocessed pictures, and did it without sending anything to the ground first.

That sounds modest and is not. The usual process is that a satellite photographs an area, waits until it passes over a ground station, transmits an enormous raw image, and then people or computers on Earth look for what matters. That chain takes hours and is limited by how much data can be squeezed through a radio link. Running the analysis in orbit inverts it: the satellite works out what is in the picture and can send down the answer — a ship here, a fire there — in minutes, at a tiny fraction of the bandwidth.

The hardware doing this is a small Nvidia processor of the kind used in robots and self-driving prototypes, hardened for space. The same chips that run AI in data centers are now, in a much smaller form, running it in orbit.

The uses are the ones you would expect and a few you might not: spotting wildfires as they start, tracking ships that have switched off their transponders, assessing flood or earthquake damage while rescue is still being organized, monitoring crops. All of these exist today and all of them are slow. Minutes instead of hours changes what they are for.

It is also a small preview of a larger argument now underway about whether serious computing belongs in space at all — where sunlight is constant and cooling is free, but everything is expensive to put there and impossible to repair.