An Earth observation satellite has, for the first time on record, found what it was looking for without human analysts in the loop. In April, Loft Orbital's Yam-9 spacecraft ran Gemma 3, Google DeepMind's vision-language model, directly on orbit and identified areas of interest from natural language queries — the first reported use of a VLM in space. The demonstration used a software harness built by NASA's Jet Propulsion Laboratory and ran on a Nvidia Jetson Orin AGX GPU onboard the satellite.
Yam-9 launched in the fall of 2025 as a pathfinder for Loft Orbital's on-orbit AI work. Researchers asked the model to classify sensor data where natural environment meets human development, and to identify infrastructure around railway hubs. It returned answers without first beaming raw imagery to the ground, where analysts typically wade through large data dumps to flag anything useful.
Gemma 3 is purpose-built for edge applications, meaning it is designed to run on constrained hardware rather than in a data center. JPL's Juan Delfa Victoria led development of NAVI-Orbital, the software package that wrapped the model for spacecraft use. Engineers had to strip down libraries and trim memory overhead to fit the orbital hardware envelope, even with one of the strongest chips currently flying in space.
Key facts
- 01Loft Orbital's Yam-9 ran Google DeepMind's Gemma 3 vision-language model in orbit in April, the first reported VLM use in space.
- 02Yam-9 launched in fall 2025 and carries an Nvidia Jetson Orin AGX GPU, one of the leading chips deployed for on-orbit compute.
- 03NASA JPL built NAVI-Orbital, the software harness that ran Gemma 3 on the satellite's constrained hardware.
- 04Loft currently operates 12 spacecraft and estimates 50 to 100 satellites would be needed for real-time global coverage.
- 05Loft recently signed a deal to build, launch and operate 6 new satellites for EarthDaily.
The near-term value is data triage. Instead of dumping terabytes of imagery to the ground, a satellite running a VLM can decide what is worth sending. The longer-term value is autonomy: a satellite that can take an instruction in plain English and act on it across days or weeks, without a roundtrip to a human operator.
Loft Orbital sells its spacecraft as platforms for third-party customers — infrastructure-as-a-service rather than bespoke satellite manufacturing. The company recently signed a deal to build, launch and operate 6 new satellites for EarthDaily, which will analyze and sell the data those spacecraft collect. Loft currently operates 12 spacecraft on orbit.
Other operators are moving in the same direction. Planet Labs flies satellites with Nvidia Jetson Orin processors and currently uses them for simpler object-detection tasks, with research underway on VLMs. Kepler Communications, which operates what it describes as the largest fleet of GPUs in space following its January launch, declined to confirm VLM deployment but cited several undisclosed use cases tied to partner NDAs.
Lasserre put a number on what full coverage would require: somewhere between 50 and 100 satellites like Yam-9 to give real-time eyes on anywhere on Earth. That is an order of magnitude beyond Loft's current 12-spacecraft fleet, and it is the constraint that will shape how aggressively the company and its customers spend on on-orbit compute over the next several years.
The same hardware lessons will inform larger ambitions. Power and memory management at orbital scale are the unglamorous bottlenecks for any future attempt to put serious compute infrastructure in space. Running a compact model like Gemma 3 on a Jetson Orin AGX is a test bed for the much harder problem of running data-center-class workloads outside the atmosphere.
NAVI-Orbital also has a second life as a building block for crewed missions. The concept began with JPL researcher Taran Cyriac John, who was thinking about digital assistants for astronauts on the Moon or Mars — operators in pressurized suits who cannot easily tap a keyboard to query a system. An interactive AI assistant, Delfa Victoria said, is the obvious workaround.
The honest caveat is that this is still a single demonstration on a single pathfinder spacecraft, using an off-the-shelf model that was streamlined to fit. The leap from one Yam-9 in orbit to a 100-satellite constellation patrolling borders, coastlines or supply chains in real time requires both a working business case and a willingness from governments and customers to let autonomous models make the first call on what counts as suspicious.
The economic implication for the Earth-observation market is the more interesting one. The value of a satellite has historically been bounded by the bandwidth available to get its data down and the human labor available to interpret it. A VLM that runs on-orbit collapses both constraints at once — and rewrites what an EarthDaily, a Planet Labs, or a defense customer is actually paying for. The asset is no longer the pixels; it is the answer.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
