NASA and IBM put an AI model in orbit — and it actually works
NASA and IBM have deployed an AI foundation model in orbit for the first time, running it on two separate platforms simultaneously. The model, called Prithvi Geospatial, was tested on Australia's Kanyini satellite and the IMAGIN-e payload aboard the International Space Station. The results confirm that flexible, general-purpose AI can operate in the constraints of space hardware — a meaningful shift from the rigid, single-task algorithms satellites have relied on for decades.
What Prithvi actually does
Prithvi was trained on a combined archive of NASA Landsat and ESA Sentinel-2 imagery spanning 13 years. That breadth lets it recognize complex patterns — flood extents, cloud cover, crop health — without being retrained from scratch for each task. The updated version, Prithvi-EO-2.0, packs 600 million parameters and scores 8% higher than its predecessor on GEO-Bench, a standard benchmark for geospatial AI, per IBM Research.
Because NASA and IBM released Prithvi openly on Hugging Face, researchers at the University of Adelaide and Australia's SmartSat CRC could adapt it for low-power satellite hardware without building a new architecture. Project lead Andrew Du said the open-source foundation cut years off the deployment timeline.
The bandwidth problem it solves
Modern Earth observation satellites generate far more data than they can transmit. Ground links are slow and expensive relative to raw sensor output, so most satellites compress aggressively or discard data before downlinking. Traditionally, that meant hardcoding narrow algorithms onboard — good at one job, useless at another.
Foundation models flip that logic. Instead of uploading new software for every new task, operators can send a small adapter module — a few megabytes — that steers the existing AI toward a different problem. The satellite stays current without a full software overhaul. NASA Chief Data and AI Officer Kevin Murphy has pointed to exactly this kind of flexibility as the reason the agency publishes its models as open source, according to NASA Science.
The real-world value is already visible. IBM used Prithvi-EO-2.0 to map the catastrophic Valencia floods of October 2024, combining Sentinel-1 radar and Sentinel-2 optical data to see through cloud cover and measure flood extent quickly after the disaster.
What comes next
NASA is not stopping at Earth observation. The agency released Surya, a heliophysics foundation model, in 2025, and is developing equivalent systems for planetary science and astrophysics. Researchers working on the current project suggest that future operators may interact with spacecraft using plain-language queries — asking a satellite what it observed the same way you'd ask a search engine.
The broader implication is straightforward: a satellite is no longer just a flying camera. With a foundation model onboard, it becomes a platform that can be repurposed, updated, and queried — without waiting for a new mission.