AI / Edge Processing in Space
Updated October 5, 2026
AI models running in orbit — from edge inference to full LLM training in space.
The bottom line
Processing is moving onto the spacecraft so raw data never has to come down — now backed by hyperscaler silicon and billion-dollar contracts.
What is happening
The traditional model — download raw data, process it on the ground — is giving way to onboard intelligence that filters and acts before anything reaches Earth. Earth observation leads: discarding cloudy frames in orbit, detecting vessels in real time, spotting wildfires in minutes rather than hours.
The constraints are unchanged and unforgiving: radiation hardening, power budgets, and the difficulty of updating a model you cannot reach. The gap between space-qualified and terrestrial processors is narrowing, which is what makes the rest possible.
Latest signals
October 1, 2026India launched its first edge-AI mission, TakeMe2Space MOI-1A, carrying NVIDIA Orin NX processors to run data processing in orbit.
Source: TakeMe2Space
September 9, 2026Loft Orbital and Marlan Space signed a $1B agreement for 50 LEO satellites running onboard models from Mistral AI.
Source: Loft Orbital
June 1, 2026Maris-Tech unveiled Venus-Space, an onboard AI video processing platform for LEO nanosatellites.
Source: Maris-Tech
April 1, 2026Meta and Booz Allen Hamilton deployed Llama 3.2 aboard the ISS as "Space Llama", validating LLM inference in orbit.
Source: Meta
Open challenges
- Radiation-hardened processors lag behind terrestrial chip capabilities
- Power budget constraints limit AI model complexity onboard
- Software update and model deployment to satellites is risky and slow
- Thermal management for compute-intensive workloads in space
- Limited ground truth data for training space-specific ML models
Companies to watch
Spire Global, Planet Labs, Google, Microsoft, NVIDIA, Ubotica, Pixxel, Starcloud, Meta, Maris-Tech, Loft Orbital, Marlan Space.