An Earth-observation satellite has just done something analysts on the ground normally handle: it found its target on its own. The demonstration, carried out on Loft Orbital’s YAM-9 spacecraft, is the first recorded use of a vision-language model in orbit and a tidy warning shot for anyone still imagining satellites as dumb cameras that simply dump data home.
Instead of sending everything down for later review, the system on board responded to natural-language prompts and identified areas of interest directly in space. That matters because the flood of raw imagery is one of the biggest bottlenecks in remote sensing, and it is exactly the sort of job AI is increasingly being pushed to do before a human ever sees the file.
How the YAM-9 satellite AI demonstration worked
The software stack was developed by NASA’s Jet Propulsion Laboratory, while the vision-language model used in the demo was Gemma 3 from Google DeepMind. The model is designed for edge use, which means it can run on constrained hardware far from a data center – a neat fit for orbit, where every watt and every byte has to earn its keep.
Researchers asked it to classify sensor data at the boundary between natural environments and human activity, and to identify infrastructure around rail hubs. It did the job, which is the point: this is not a flashy chatbot in space, but a machine that can look at imagery and answer a specific question without waiting for Earth to chime in.
Why onboard processing changes satellite value
Short term, this could make satellites more useful by filtering data before it ever leaves orbit, cutting down the stream of unprocessed imagery that teams must sort through. That is a very practical win, and a very unglamorous one – which usually means it is the kind of thing that scales.
Longer term, it points to something bigger: the possibility of building heavier AI infrastructure in space itself. The space industry has already spent years moving from brute-force data collection toward smarter, software-driven operations, and this demo suggests the next step is not just better sensors, but sensors that can think a little before they talk.
The next step for space-based AI
Loft’s head of AI said the approach could support continuous patrol-style monitoring, where a satellite is effectively told to watch a border or site and flag anything suspicious. NASA’s own AI lead framed the same idea from a crewed-spaceflight angle: astronauts, like most humans, do not need more dashboards, they need help with complex tasks.
The real question now is not whether a satellite can run a model. It is how far operators will trust onboard AI to decide what matters, and how quickly commercial and government missions will stop treating orbit as a relay station and start treating it like a place where computation actually happens.