Planet Labs runs AI in orbit
Planet Labs put Nvidia Jetson Orin modules on its Pelican spacecraft to run on-board AI that detects planes from satellite imagery within minutes of capture, rather than waiting for ground processing (x.com). That approach reduces the data-to-insight lag for time-sensitive imagery and shows how pushing inference to edge hardware changes satellite ops tradeoffs (x.com).
A satellite usually works like a camera trap in the woods: it snaps the picture in orbit, sends the whole file down to Earth, and only then does a computer decide what was in it. Planet Labs just showed a different model, with the computer riding on the spacecraft itself. (businesswire.com) On March 25, 2026, Planet’s Pelican-4 satellite flew about 500 kilometers above Alice Springs, Australia, captured an airport image, and ran an artificial intelligence model onboard to find airplanes “in moments.” Planet said the hardware doing that work was an Nvidia Jetson Orin module mounted inside the satellite. (businesswire.com) That onboard model is doing inference, which is the part where a trained system looks at a new image and makes a call, like a cashier scanning items instead of restocking the whole store. The important shift is that Pelican-4 did the recognition step before the image ever waited in line for a ground station. (nvidianews.nvidia.com) (businesswire.com) Why that delay matters: satellites can collect data continuously, but they can only dump data when they have a communications link, and high-resolution imagery is heavy. Planet says moving analysis to the edge can cut the path from “seeing” to “acting” from hours to minutes while also reducing downlink latency and cost. (planet.com) (businesswire.com) Pelican is Planet’s newer high-resolution line, built for faster tasking and multiple revisits per day, with first-generation imagery listed at 50-centimeter resolution on Planet’s product page. The company says the satellites are aimed at fleeting events, the kind that can change before the next download window opens. (planet.com) Planet launched Pelican-3 and Pelican-4 in August 2025, calling them “AI-enabled” when they went up, so this week’s news is the first public proof of what that extra compute was for. In other words, the chip was not just along for the ride; it was part of the satellite design from the start. (businesswire.com 1) (businesswire.com 2) The first result was not perfect. Planet said the initial airplane detection hit 80 percent accuracy on raw imagery, and the company is still working on precision and recall, which are the two knobs that control false alarms and missed targets. (businesswire.com) Nvidia has been pitching this exact use case since March 16, saying Jetson Orin is meant for size-, weight-, and power-constrained missions where a full data center cannot fit. Spacecraft care about every watt and every gram, so a compact edge computer changes what can be done before data ever touches the ground. (nvidianews.nvidia.com) Planet is framing this as the start of a near-real-time network across its Pelican satellites and its planned Owl constellation, tied together with high-speed inter-satellite links. If that buildout works, the satellite stops being a flying camera and starts acting more like a sensor with judgment. (businesswire.com) The practical tradeoff is simple: sending every pixel home gives analysts maximum flexibility, but it burns bandwidth and time; deciding something in orbit is faster, but now the model onboard has to be trusted. Planet just showed where the industry is headed: fewer raw images waiting for Earth, more answers arriving with the image. (planet.com) (businesswire.com)
