Google's Gemma AI Has Run in Orbit to Decide What Data Is Worth Sending Down
In a small test, only 11.7% of the data was worth the $10 to $20 per gigabyte trip to Earth.

This Week in Startups · Running Google’s Gemma on Satlyt’s orbital AI (w/ CEO Rama Afullo)Satlyt, a company that writes software to coordinate work across satellites, has run Google's Gemma, a small artificial intelligence model built to run on modest computers, aboard satellites in orbit, chief executive Rama Afullo says. Google DeepMind has published a case study on the experiment.
The test shows how satellites could sort data in space before paying to send it home. That matters because getting data to Earth through ground stations can cost $10 to $20 per gigabyte.
The figure comes from a small demonstration already flown on a couple of satellites. In it, a satellite could not initially send data to its ground station, possibly because of cloud cover or electromagnetic interference.
The data was passed and relayed toward the ground station, while the onboard AI diagnosed the problem and shrank the file so it was small enough to send down. In that demonstration, Afullo said, only 11.7% of the data was useful enough to send to Earth, and sending only what was useful was cheaper and more efficient.
Afullo said NASA has tasked Satlyt with using the same relay approach for a lunar base. The idea is to carry high-resolution video from the lunar surface through orbiters circling the Moon, then hand it down through satellites nearer Earth.
Questions people ask
What is Gemma?
Gemma is Google's family of lightweight, openly available AI models. They are much smaller than leading frontier models, so they can run on a laptop, a phone, or in this case, a computer aboard a satellite.
Why does it cost money to send data down from a satellite?
Satellite operators generally do not own antennas all over the world, so they rent time from ground-station networks. Those operators charge for the antenna time, the bandwidth, and the handling, which is commonly billed by the amount of data delivered.
Why not just send everything down and sort it on Earth?
Bandwidth from space is limited, links are intermittent, and a satellite may pass over a ground station only briefly. Sorting overhead means less waiting, lower bills, and faster delivery of the useful part, especially for time-sensitive uses like weather or emergency information.
