Loft Orbital's Yam 9 used a Google DeepMind vision language model on orbit in April 2026, the first reported autonomous find by an Earth observation satellite, and a test of whether spacecraft can stop shipping raw pixels home.
In April 2026, an Earth-observation satellite called Yam-9 did something its predecessors cannot. It ran a vision-language model on orbit, accepted a plain-English prompt, and returned a map of railway-hub infrastructure and the edges where natural land meets human development, all without waiting for a ground analyst to sift through raw imagery.
The shift is architectural. For decades, an Earth-observation satellite has been a dumb sensor in the sky. It collects pixels, dumps them to a ground station, and waits for a human to decide what matters. The bottleneck has been downlink bandwidth and analyst time, not sensor quality. Loft's head of AI, Paul Lasserre, told TechCrunch the on-orbit model "opens the door to always-on, patrol layers in space… monitor this border for me, and let me know when something is suspicious." That is a sales pitch, and the underlying capability is real, but the economics are just starting to point in the same direction.
The hardware is a standard commercial edge-AI kit. Yam-9 carries an NVIDIA Jetson Orin AGX GPU, the same class of chip that powers autonomous drones and robotaxis on the ground. The software is a vision-language model, Gemma 3, released by Google DeepMind as a small, edge-targeted model. Sitting on top is NASA JPL's NAVI-Orbital, a system built by JPL technical leader Juan Delfa Victoria to make models usable under the radiation, thermal, and power constraints of low Earth orbit. Engineers had to slim Gemma 3 down to fit the limited on-orbit memory and library footprint, a reminder that space-grade inference is not the same problem as cloud inference.
The tasks the pathfinder has actually demonstrated are narrow classification. Areas where natural environment meets human development. Infrastructure around railway hubs. The kind of questions a geospatial analyst would otherwise queue up in a notebook. The model is not reasoning about geopolitics. It is doing fast triage so that the high-resolution pixels that do get beamed home are the ones a customer is likely to pay for. For Loft, that is the path from selling pictures to selling answers, and it is the bet behind a recent deal with EarthDaily to build, launch, and operate six new satellites with onboard data analysis. Loft currently flies 12 spacecraft and talks about a future constellation of 50 to 100 for real-time global coverage, though that target is a company aspiration, not a committed deployment plan.
The neighbors are moving in the same direction. Planet Labs already flies Jetson Orin hardware on its birds and is running simpler object detection on orbit, with VLM research underway. Kepler Communications launched what it calls the largest orbital GPU cluster in January 2026 and has signaled that it has VLM-class workloads running for customers under NDA. The competitive question is no longer which company has the best camera. It is which company can afford to put the GPU cluster in orbit, and which business model can pay for the inference cost before the bytes ever leave the spacecraft.
The honest limit is that this is still a single pathfinder. Yam-9 launched in fall 2025 specifically to test whether this kind of inference survives launch, radiation, and the thermal swings of low Earth orbit. It has, at least for narrow tasks, on a single bus. The originating idea behind the software, called NAVI-Space, came from JPL researcher Taran Cyriac John as a digital assistant for suited astronauts on the Moon and Mars. The Yam-9 demo is a terrestrial proxy for that long-horizon goal. Whether the same approach scales to dozens or hundreds of satellites, and whether the on-orbit inferences stay correct under sustained radiation and limited power, are the open engineering questions that will decide whether on-orbit AI becomes infrastructure or stays a marketing line.