The satellite industry just figured out how to think before it talks.
The Summary
- Planet Labs successfully ran AI object detection on a satellite in orbit, marking the first reliable onboard image processing for Earth observation—a capability the industry has chased for years.
- The breakthrough eliminates 6-12 hour delays between image capture and actionable intelligence by processing data in space rather than beaming raw files to ground stations.
- Planet's constellation generates 30 terabytes daily (10,000 hours of HD video equivalent), which previously required hours of ground transfer and cloud processing before AI analysis could even begin.
The Signal
Planet Labs' Pelican-4 satellite identified aircraft on an Australian tarmac and drew green boxes around them—all without phoning home first. That's the engineering flex, but the economics are what matter. The company operates hundreds of satellites producing image data faster than their ground infrastructure can meaningfully process it. The 6-12 hour lag between capture and insight meant they were running the world's most expensive historical archive, not an intelligence service.
Eighteen months of engineering work went into making AI models reliable enough to run on satellite hardware. Not groundbreaking models, just stable ones. The constraints are brutal: limited compute, radiation exposure, no software updates you can roll back if something breaks. Vice president Kiruthika Devaraj frames it plainly—Planet has "very good eyes in space looking at everything," but those eyes were essentially blind until the data hit Earth.
"We collect so much data and have to wait six to 12 hours to get the information out. So, you're essentially looking at the past."
The shift from dumb sensors to smart satellites changes what you can sell. Raw imagery is a commodity. Real-time intelligence about what's actually happening—a wildfire sparking, ships changing course, construction projects appearing overnight—that's a service worth paying for. Planet's constellation includes hundreds of Dove cubesats (30 cm long, 5-meter resolution) and 32 larger Pelicans (30 cm resolution). The fourth Pelican, launched in 2025, ran the airport detection demo.
Key capability unlocks:
- Autonomous satellite tasking: satellites decide what's worth imaging next without ground control latency
- Bandwidth optimization: transmit alerts and insights, not terabytes of raw pixels
- Event detection speed: wildfires, military movements, or infrastructure changes flagged in minutes, not half a day later
The data volume problem isn't going away. 30 terabytes per day flows through tens of ground stations worldwide, gets transferred to cloud storage, then finally reaches AI processing pipelines. Every step burns hours. Moving inference to orbit doesn't just speed things up—it flips the entire data architecture. Instead of "collect everything, sort it out later," you get "understand in real time, transmit what matters."
This matters beyond Earth observation. Every sensor-heavy infrastructure faces the same problem: too much data, not enough intelligence at the edge. Autonomous vehicles can't wait for cloud decisions. Factory sensors can't batch-process quality control. The pattern is consistent—push inference to where the data originates, or drown in latency and transfer costs.
The Implication
The satellite operators who figure out onboard AI first will redefine what Earth observation sells. Not pictures. Intelligence. Planet Labs spent 18 months on airplane detection because getting AI stable in orbit is harder than getting it accurate on the ground. But now the playbook exists. Expect every serious constellation to follow.
For anyone building agent systems that need real-world context—logistics, insurance, agriculture, defense—the quality of your inputs just improved and the latency just dropped. The planet is becoming more legible, faster. Build accordingly.