The bottleneck between observation and response has been shown removable.
For generations, the gap between seeing a disaster from space and responding to it on the ground has been measured in hours — hours that cost lives. NASA and IBM's Prithvi AI model, trained on thirteen years of Earth observation, has now crossed into orbit aboard two active spacecraft, performing the work of analysis in space itself rather than waiting for data to descend. The deployment, a collaboration spanning universities, space agencies, and governments across three continents, represents the first time a large-scale foundation AI model has operated successfully in orbit — and with it, the long delay between observation and understanding has been shown, for the first time, to be a problem we can solve.
- Every hour lost between a satellite spotting a flood and a responder receiving actionable data is an hour in which the disaster deepens — Prithvi was built to collapse that window.
- The model is now running aboard the South Australian government's Kanyini satellite and the ISS, having already detected floods, wildfires, burn scars, and crop stress in early tests.
- A foundational technical barrier has been cleared: no large general-purpose AI had ever been successfully operated in orbit before this deployment.
- Prithvi's modular architecture allows new disaster-detection capabilities to be added via small software packages, meaning satellites can evolve without being rebuilt or replaced.
- Open-source availability was decisive — without it, the lead researcher estimates the project would have required years of additional work just to build a comparable starting point.
- The demonstration signals a near future in which Earth-monitoring satellites launch with onboard intelligence already installed, making real-time disaster response a global baseline rather than an exception.
For the first time, an artificial intelligence model trained to read Earth's surface is doing that work not on the ground, but in orbit. NASA and IBM's Prithvi — built from thirteen years of global satellite imagery — is now running aboard two active spacecraft: the Kanyini satellite above South Australia and a payload docked to the International Space Station. In early tests, it has already detected floods, traced wildfire burn scars, and monitored crop health.
The stakes of this shift are rooted in a simple inefficiency. When a satellite currently observes a disaster, raw data travels to Earth, waits in an analysis queue, and returns as actionable intelligence — a process that can take hours. Prithvi performs that analysis while still in orbit, compressing the gap between observation and understanding from hours to minutes. For emergency responders, that compression is the difference between a map of where a flood was and a map of where it is.
The project brought together Adelaide University, the European Space Agency's Φ-lab, Thales Alenia Space, and Australia's SmartSat Cooperative Research Centre. A preprint confirms the model performed as intended across both platforms. What made it practical is Prithvi's architecture: rather than requiring costly full software uploads, new capabilities can be delivered as small decoder packages — meaning a satellite could gain the ability to detect a new type of disaster without being redesigned.
Dr. Andrew Du of Adelaide University, who led the project, credited the model's open-source availability as essential. Without it, he noted, building a comparable foundation would have consumed years. NASA's chief science data officer framed the deployment as proof of a broader principle: open tools accelerate progress beyond any single mission. The bottleneck between space observation and human response has been shown to be removable. The work now is making that removal routine.
For the first time, an artificial intelligence model trained to read Earth's surface has left the ground and started working in the vacuum. NASA and IBM's Prithvi, built on thirteen years of satellite imagery, is now running aboard two active spacecraft—one orbiting above South Australia, the other docked to the International Space Station. In early tests, it has already done what it was designed to do: spot floods spreading across land, trace the charred edges of wildfires, and watch crops grow.
The significance of this moment sits in a simple fact about how we currently monitor disasters. When a satellite observes a flood or fire, the raw data travels down to Earth, where it waits in a queue to be analyzed. Hours pass. By the time a useful picture emerges, the situation on the ground has already shifted. Prithvi changes that equation by doing the thinking in space. The model identifies patterns while the data is still in orbit, compressing the time between observation and understanding from hours to minutes. For emergency responders, that gap is everything.
The deployment involved researchers from Adelaide University, the European Space Agency's Φ-lab, Thales Alenia Space, and Australia's SmartSat Cooperative Research Centre. They uploaded Prithvi to the Kanyini satellite, operated by the South Australian government, and to the IMAGIN-e payload aboard the ISS. A preprint of their work confirms the model performed as intended, detecting floods and clouds across both platforms while also mapping burn scars and monitoring crop health. The fact that it worked at all represents a technical hurdle cleared: no one had successfully run a foundation model—a large, general-purpose AI system—in orbit before.
What makes Prithvi practical for space is its architecture. Traditional satellites cannot easily accept large software updates because bandwidth in orbit is scarce and expensive. Prithvi solves this by functioning as a foundation model, meaning new capabilities can be added through small decoder packages rather than uploading an entirely new system. A satellite that was never designed to detect a particular type of disaster could, in theory, gain that ability without being redesigned or replaced. It is a form of flexibility that space hardware has never really had.
Dr. Andrew Du, a postdoctoral researcher at Adelaide University and AI engineer at SmartSat, led the project. He emphasized that none of this would have been possible without Prithvi being open source. "If Prithvi weren't open source, I would have to train my own foundation model," Du said. "Having that model openly available saved a lot of time and effort." The model was trained by IBM and NASA's IMPACT AI team at Marshall Space Flight Center using the Harmonized Landsat and Sentinel-2 dataset, a combined archive of more than a decade of global satellite imagery from NASA and the European Space Agency.
Kevin Murphy, NASA's chief science data officer, framed the deployment as validation of a larger principle. By making tools like Prithvi available to anyone who wants to use them, he argued, the agency accelerates scientific and technological progress. The implication is clear: this is not the end of the story. Other researchers, other space agencies, other countries can now build on what has been demonstrated. The next generation of Earth-monitoring satellites could launch with Prithvi already aboard, or with the ability to receive it. The bottleneck that has always existed between observation and response—the time it takes to move data from space to ground to analysis—has been shown to be removable. What comes next is the work of making that removal routine.
Notable Quotes
If Prithvi weren't open source, I would have to train my own foundation model. Having that model openly available saved a lot of time and effort.— Dr. Andrew Du, Adelaide University and SmartSat
By sharing these tools with anyone who wants to use them, we accelerate scientific and technological development into the future.— Kevin Murphy, NASA Chief Science Data Officer
The Hearth Conversation Another angle on the story
Why does it matter that the analysis happens in space instead of on the ground? Isn't it just a matter of speed?
Speed is part of it, but it's more fundamental than that. When a flood is happening, every hour of delay means more people at risk, more infrastructure damaged. If you can identify the flood while the satellite is still overhead, you can alert responders immediately. On the ground, you're always chasing what already happened.
But couldn't you just build faster ground systems? Why send the AI into orbit?
You could, but there's a hard limit. The data has to travel from orbit to Earth first—that's physics. Even at light speed, there's latency. And then it has to be processed, interpreted, transmitted to the people who need it. In orbit, Prithvi sees the flood and knows what it is before the data even starts its journey down.
The article mentions this is a foundation model. What does that actually enable that a regular satellite AI couldn't?
A foundation model is like a general-purpose brain. You train it on a huge range of Earth observation data, and then you can teach it new tasks without retraining from scratch. You just send a small decoder—a lightweight instruction set—and suddenly it can do something new. A satellite built ten years ago could gain capabilities its designers never imagined.
So the open-source part—that's just about cost savings for researchers?
It's bigger than that. It's about democratization. If Prithvi were proprietary, only the organizations that built it could use it. By making it open, they're saying: anyone with a satellite can now have this capability. A developing country with limited resources could deploy it. A university could experiment with it. That's how you get innovation you can't predict.
What happens if the model makes a mistake in orbit? Can they fix it?
That's where the architecture matters again. If there's a flaw in how it's identifying floods, they can send up a corrected decoder without replacing the whole system. It's not perfect—there will always be edge cases—but it's far more resilient than the old way of doing things.