Satellite AI Shifts to Orbit as Foundation Models Run Onboard

This year's coverage of AI in satellite operations often comes as listicles that rank "critical solutions" for managing spacecraft. The reporting behind those lists tells a more specific story. The main change in 2026 is not one product. AI models are moving out of ground-station software and onto the satellites. These include general-purpose foundation models, reinforcement-learning controllers and agent-based pipelines. The ground segment, meanwhile, is being rebuilt around automation that operators say they can no longer do without.

The research news points the same way. Benchmarks and assurance methods are struggling to keep up.

Foundation models reach orbit

The clearest research milestone of the year came in May. NASA reported that its Prithvi geospatial model, built with IBM, became the first geospatial foundation model deployed in orbit.13 Researchers from Adelaide University, ESA's Φ-lab, Thales Alenia Space and South Australia's SmartSat Cooperative Research Center uploaded a compressed version of the model to two platforms. One was South Australia's Kanyini satellite. The other was Thales Alenia Space's IMAGIN-e payload on the International Space Station. On both, they tested how well the model detected floods and clouds.13

The engineering detail matters for satellite management. Once a foundation model is on board, operators do not need to upload a whole new model to give it a new task. They can send a small decoder package, which uses much less bandwidth.13 For spacecraft with limited uplink time, that turns software updates into a lightweight operation.

Prithvi was not the only model in orbit. One industry newsletter reported that IBM's TerraMind.tiny now runs on Planetek Italia's AI-eXpress constellation. It described this as an early test of whether general-purpose models can replace task-specific algorithms on satellites.16 The same report said hardware supply is no longer the bottleneck. Ubotica has flown 11 missions with more than 30 AI models in orbit. Other companies, including Unibap, KP Labs, EDGX and Aethero, are shipping compute hardware.16

On the ground, new models keep coming. ESA reported wider availability of Tessera in June 2026. Tessera is a foundation model trained on Sentinel-1 radar and Sentinel-2 optical data.12 In July, Google opened a private preview of Custom Satellite Embeddings, which summarizes observations for places and time periods that users choose.18 In August, LiveEO received a high six-figure grant from the German Space Agency to study a multimodal Vision Foundation Model. The model would detect change across optical and radar imagery within a single architecture.17

The benchmark problem

The trend reports mostly leave out an important caveat. Independent evaluations do not show that foundation models win across the board. The PANGAEA benchmark found that foundation models do not consistently outperform supervised alternatives.18 GEO-Bench-2 tests models on 19 datasets, and its authors reported that no single model dominates every task.18 General-purpose image models did well on high-resolution tasks. Models built specifically for Earth data did better on multispectral applications.18

This tension runs through the year's coverage. Some trend pieces predict that EO foundation models will become standard infrastructure, comparable to LLMs, within five years.14 More careful analysis says the savings depend on the task, and that adapting and validating the models still costs money.18 Our reading is that both views can be true. Foundation models will probably become the default starting point, but they will not remove the need for checks specific to each task. The in-orbit demonstrations so far cover a narrow set of tasks, such as flood and cloud detection, rather than full mission autonomy.13

Collision avoidance becomes a learning problem

The second research front is space traffic. Here the main development is that onboard controllers are starting to rely on learned methods instead of fixed rules. A patent survey describes two 2025 Beihang University filings. They cover neural-network collision avoidance that runs on limited onboard processors and does not depend on commands from the ground. A 2025 US patent uses deep reinforcement learning for onboard trajectory control.21 The same survey says the field is settling on a hybrid design. Ground systems would handle strategic conjunction planning, and onboard AI would handle emergencies.21

Academic work is testing newer architectures. Orbit-Planner, submitted to arXiv in August 2026, is a latent world model for avoiding obstacles in orbit. It learns how a spacecraft moves in response to its actions and simulates future states. In closed-loop simulation in Isaac Sim, it succeeded 91.7% of the time.27 A May 2026 paper proposed a pipeline with separate agents for scouting, analysis, planning, safety and operations. It reported 97.23% classification accuracy across four risk tiers.24 Earlier reinforcement-learning research framed collision avoidance as a decision problem under partial observation, where the controller tries to minimize both collision risk and fuel use.25

The sensing side is advancing too. In September, researchers at the Alan Turing Institute's Defence AI Research Centre published an anomaly-detection model. It learns from "light curves," which are readings of how much light satellites reflect.22 It identified unusual light curves 88% of the time and could tell a spinning satellite from a tumbling one. It flags anomalies for human experts to investigate rather than acting on its own.22

These results should be read with care. A 91.7% success rate in simulation and a 97% score on a held-out test set are research results. They do not show the models are ready to fly. Even so, the direction is clear. Learned controllers are being designed for the computing limits of real spacecraft, not just for ground servers.

Operations: how much autonomy is real?

The coverage disagrees most on the question of autonomy. One analysis of the ground segment says current practice still involves heavy human supervision. In that view, AI recommends a fix or a maneuver, and a human approves it. Starlink is an exception: it already lets satellites dodge on their own in time-critical cases.1 Startup marketing goes further. Constellation Space, backed by Y Combinator, says its ConstellationOS predicts link failures with more than 90% accuracy and reroutes traffic in under two seconds without human involvement.2 At a September industry panel, executives described humans as moving into supervisory roles. They also stressed that people are still needed to validate AI decisions, train models and do context engineering.8

We think the panel's view is the most realistic. Full autonomy is real today only in narrow areas where speed matters most, such as collision dodging and network rerouting. Elsewhere, operators are moving toward a model where AI acts within set limits and passes unusual cases to humans. One analysis describes exactly that next step: AI handles routine actions within defined boundaries and escalates the rest.1

The economic case for automation is consistent across outlets. One analysis calls automation a structural requirement for running large constellations at commercial cost.1 Market estimates put AI in space operations at about $2.89 billion in 2026, growing about 22.9% a year.41 Not every number agrees. Kratos cited more than 9,000 Starlink satellites in late December 2025.5 Another analysis put the figure above 10,000 by spring 2026.1 That difference likely reflects launches in between rather than an error in either count.

Hardware and agents shape the next model generation

Hardware news shows where model design is heading. Nvidia used GTC 2026 to announce its Vera Rubin Space-1 Module for running inference in orbit. Planet Labs already uses Nvidia IGX Thor processors to classify images on board and send down change-detection results instead of raw pixels.1 Planet's next Owl constellation is designed with onboard GPUs from the start. Satellogic's "AI-First" design runs several models at once that can be updated after launch.16

Research also shapes how these models are built and checked. Advanced Space, working with NASA, trains neural networks for station-keeping that follow physical laws by building conservation principles into training.5 An ESA-funded project by Craft Prospect and GMV pairs reinforcement learning for mission planning with transformer models for anomaly detection. A verifiable onboard supervisor checks the AI's plans.3 That design matters. It suggests the field's answer to trusting AI is to wrap it in deterministic checks, not to rely on explainability alone.

Agent-style AI is also arriving. Speakers at the September panel said agentic AI and new foundation models are enabling more end-to-end, autonomous work, with AI defining the actions a satellite takes.8 Slingshot Aerospace's TALOS system, deployed to the US Space Force, uses AI agents to simulate realistic spacecraft threats for training.1

Governance trails the models

Regulation is the gap the coverage most often points to. One overview notes that no international treaty covers autonomous satellite decisions. It also says there is no shared protocol for coordinating collision avoidance between operators.4 Another analysis expects the FCC, the Space Force and the Office of Space Commerce will need to define what operators may let satellites do without human review.1 NASA's coordination with Starlink gives an early example: humans set the rules of engagement, written with autonomous systems in mind.7

Our reading is that 2026's real progress is the new generation of models, not a list of products. Foundation models now run in orbit, learned controllers are being built for real spacecraft hardware, and agent pipelines are starting to handle conjunction screening. Independent benchmarks, verified autonomy limits and rules shared across operators are still missing. That will decide whether these models move from demonstrations to trusted infrastructure.