Loft Orbital YAM-9 Satellite Deploys Gemma 3 AI Onboard
Key insights
- YAM-9 became the first Earth observation satellite to autonomously classify objects using AI, running Gemma 3 on Nvidia Jetson Orin AGX hardware in April 2026.
- NASA JPL's NAVI-Orbital software translates natural-language queries into onboard satellite image classification, eliminating the need for raw data downloads to Earth.
- Loft Orbital operates 12 satellites today and estimates 50-100 spacecraft would provide real-time always-on global coverage.
Why this matters
Onboard AI inference removes the bandwidth bottleneck that has historically constrained Earth observation to post-hoc ground analysis, opening time-sensitive use cases like disaster response and persistent infrastructure monitoring. The pairing of a commercial VLM (Gemma 3) with JPL's NAVI-Orbital software shows that space-qualified AI stacks can now be assembled from existing research and consumer hardware rather than custom silicon. With Planet Labs and Kepler Communications pursuing parallel programs, the industry is moving toward a competitive market for query-driven satellite intelligence rather than raw imagery subscriptions.
Summary
Loft Orbital's YAM-9 became the first Earth observation satellite to autonomously identify objects of interest in April 2026, running Google DeepMind's Gemma 3 on an Nvidia Jetson Orin AGX processor onboard.
NASA JPL's NAVI-Orbital software answers natural-language queries like 'infrastructure around railway hubs' by classifying live imagery on the satellite itself, bypassing raw data downloads to ground analysts entirely.
Essentially: (Loft Orbital, NASA JPL, Google DeepMind) have validated the first onboard AI triage layer for Earth observation satellites.
- Loft currently operates 12 satellites and estimates 50-100 would enable real-time global coverage.
- Paul Lasserre, Loft's head of AI, described the goal as 'always-on, patrol layers in space.'
- Planet Labs and Kepler Communications are pursuing parallel programs; Kepler declined specifics, citing NDAs.
The shift reframes satellites from passive data pipes into autonomous, query-driven sensors.
Potential risks and opportunities
Risks
- Onboard classification errors could reach downstream users (disaster response, infrastructure monitoring) without the ground-truth verification that current download-first workflows provide.
- Loft Orbital's 12-satellite fleet is far short of the 50-100 needed for real-time global coverage, making the 'always-on patrol' use case dependent on significant future capital raises or customer commitments.
- Competing programs from Planet Labs and Kepler Communications could fragment the nascent market before interoperability standards for natural-language satellite query interfaces are established.
Opportunities
- Nvidia gains a new space-specific vertical for the Jetson Orin line, with potentially high-margin, low-volume contracts as more satellite operators validate onboard inference.
- Defense and intelligence customers seeking persistent area monitoring without heavy downlink bandwidth represent an immediate commercial pipeline for Loft Orbital's query-driven capability.
- NASA JPL's NAVI-Orbital software, already validated on YAM-9, positions JPL as a licensing or partnership target for commercial operators building onboard AI stacks.
What we don't know yet
- Whether Gemma 3 was fine-tuned on satellite imagery or deployed off-the-shelf; no accuracy benchmarks against ground-based analysis were disclosed.
- Kepler Communications' program details remain under NDA as of June 2026, leaving its timeline, hardware configuration, and capability scope unknown.
- No power consumption or thermal performance figures were reported for the Nvidia Jetson Orin AGX operating in the orbital radiation environment.
Originally reported by techcrunch.com
Read the original article →Original headline: Loft Orbital's YAM-9 Satellite Runs Google Gemma 3 in Orbit — First Vision-Language Model Deployed in Space Enables Natural-Language Queries Over Live Earth Imagery