High above Alice Springs, Australia, the Pelican-4 satellite sweeps across the Outback and captures an airport tarmac below. On the image that materializes inside the spacecraft's computer: thirteen aircraft, each outlined in crisp green bounding boxes. The boxes are not drawn by an analyst on the ground. They are generated by an AI model running on the satellite itself—classification happening in orbit, in real-time, with no link to Earth required.
This single image represents something the Earth observation industry has pursued for years and failed to achieve until now. Planet Labs, the San Francisco-based company that operates the world's largest fleet of commercial imaging satellites, demonstrated autonomous on-board AI inference aboard Pelican-4 on April 29, 2026. After 18 months of development, the company's engineers achieved what they call a fundamental shift: moving AI classification off the ground and onto the spacecraft where the data is collected.
The practical stakes are significant. Planet currently operates several hundred Dove and SuperDove CubeSats that scan the entire Earth daily at roughly 5-meter resolution, generating 30 terabytes of data per day. This imagery streams to ground stations worldwide, then travels to cloud infrastructure for processing—a pipeline that introduces a 6-12 hour delay between collection and actionable insight. A wildfire spotted from orbit might not reach emergency managers until it has grown beyond easy containment.
"Minutes matter in some sectors," said Kiruthika Devaraj, Planet's vice president of engineering. "Real-time insights really enable us to provide answers to problems as they're unfolding."
The technical breakthrough lies in what sits inside Pelican-4's chassis. Planet outfitted the satellite with Nvidia Jetson Orin GPU modules—the same processors that power autonomous drones and ground vehicles. The AI model analyzes a single Pelican image, comprising 16,000 pixels, in half a second. Results can reach end users within minutes of capture, compared to the half-day wait currently standard across the industry.
The Alice Springs demonstration is not a laboratory proof-of-concept. It is a live system operating in production conditions. Pelican-4, the fourth satellite in Planet's planned 32-satellite Pelican constellation, executed the inference while in low Earth orbit, proving that edge AI can function reliably outside laboratory environments.
Planet plans to extend this capability across its fleet. The SuperDove constellation will eventually carry Jetson-class processors, and a new satellite design called Owl will provide daily revisits at up to 1-meter resolution with onboard AI detection. The company frames this as a move toward what Deveraj calls "planetary intelligence"—a satellite network that makes decisions faster than any ground-based system can respond.
The implications extend beyond Planet's own constellation. If edge AI inference proves reliable on orbit, the traditional model of capturing everything and processing on Earth becomes optional rather than mandatory. The bottleneck shifts from data transmission to observation strategy—who decides what the satellites look at in the first place.
The Alice Springs image remains a single data point. But for the first time, that data point was generated entirely in space.