NASA Ran a Google AI Model Inside a Satellite. Here Is What It Found.
A small, efficient AI model analyzed real images from orbit and sent back plain-text descriptions instead of raw data. It could cut wildfire alerts from 90 minutes to near-instant.
Key points
- NASA's Jet Propulsion Laboratory ran Google's Gemma 3 AI model aboard the YAM-9 satellite in 2025, marking the first time a vision-language AI processed images in orbit from its own camera.
- The system, called NAVI-Orbital, scored 88 percent accuracy classifying 7,960 images in a ground-based test before launch.
- The model runs on just 8 GB of memory, small enough for a laptop, yet powerful enough to describe satellite imagery from space.
- Bandwidth limits currently delay wildfire satellite alerts by up to 90 minutes; text summaries from an onboard AI could cut that close to zero.
- Scientists can now direct the satellite using plain-English prompts rather than writing complex structured commands through an operations team.
A Google AI model is orbiting Earth right now, looking out a satellite window and writing down what it sees.
NASA's Jet Propulsion Laboratory (JPL) recently flew Google's Gemma 3, an open-source large language model (a type of AI that reads and generates text and, in this case, understands images too), aboard a commercial satellite called YAM-9, built by Loft Orbital. Reported by IEEE Spectrum AI, it marks the first time a vision-language AI has analyzed images from a satellite's own camera while still in orbit.
What did the AI actually do up there?
The system looked at photos taken by the satellite's camera and wrote plain-English descriptions of them. No raw image files were beamed back to Earth. Just words.
The software framework is called NAVI-Orbital. During two live tests, the satellite captured images over Toulouse, France and over the coast of Argentina. The AI studied each image and answered scripted questions about whether areas were commercial or residential, and whether natural features were visible.
All that processing happened on a small Nvidia Jetson Orin AGX, a compute module about the size of a thick paperback, commonly used in robotics. The version of Gemma 3 on board uses a compressed format needing only 8 GB of memory. Juan Delfa, the technical group lead at NASA JPL who led the project, put it plainly: "You can run it in a tiny, tiny computer."
| Detail | Figure |
|---|---|
| Ground-test accuracy | 88% across 7,960 images |
| Model size (memory) | 8 GB |
| On-board power available | 150 to 500 watts (solar) |
| Current wildfire alert delay | Up to 90 minutes |
| Data saved vs. Raw images | Kilobytes vs. Hundreds of megabytes |
Why does it matter that the AI is on the satellite rather than on the ground?
Satellites can only beam data to Earth at set intervals, and the connection is often slow. Sending a full high-resolution image consumes significant bandwidth; a two-sentence text description takes almost nothing.
Loft Orbital's general manager Paul Lasserre calls this "semantic compression": the AI summarizes what matters, and only that travels home. "It doesn't matter if the link is slow, because you're downlinking dozens of kilobytes instead of dozens or hundreds of megabytes," Lasserre said.
The wildfire example sharpens that point. We covered the new generation of fire-spotting spacecraft in our 17 July story on purpose-built microsatellites reaching orbit, and this experiment speaks directly to why speed matters: processing delays currently push alerts back by up to 90 minutes. An onboard AI flagging a fire in plain text could collapse that window.
What changes for scientists who use satellites?
Directing a satellite today means writing precise, structured commands through a specialist operations team. NAVI-Orbital lets a researcher type a plain-English instruction instead, upload it, and let the satellite act on it.
"Now, a scientist can write a prompt, upload it to the spacecraft, and that will be taken into account by the system," Delfa said. Switching from urban mapping to wildfire detection means editing a text prompt. No new software, no retraining the AI.
The system is deliberately walled off from flight controls for now. It reads images and returns descriptions, nothing more. But Delfa's longer vision is larger: he wants AI to serve as a natural-language companion for astronauts, helping them operate equipment inside a spacesuit where dexterity is limited.
That future needs years of research. What NAVI-Orbital proved is that both foundations, running a capable AI model in space and controlling it with plain language, work today.
Takeaway: The faster-alerts story is real and closer than most space technology tends to be. The piece worth watching is whether "semantic compression" becomes standard kit on Earth-observation satellites, because if it does, emergency response systems get a genuine upgrade without waiting for better bandwidth infrastructure.
