Satellites produce far more data than they can ever downlink. Sparse contact windows delay insight until it's stale — which is why reasoning keeps moving onboard, to where the data is born. The catch: AI agents want power that a satellite doesn't reliably have.
The tradeoffs aren't intuitive.
Agents live inside hard mission constraints — bounded latency, sufficient confidence, and staying alive without draining the platform. Energy availability swings with orbit and eclipse. And the model choices interact strangely: a larger model draws more instantaneous power but can finish faster, sometimes costing less total energy per query; a smaller one draws less at peak but pays in latency and accuracy. No single model configuration survives every operating condition.
Serving, not selecting.
So we stop picking a model and start serving a portfolio. The satellite carries several models with distinct accuracy, latency, and power profiles; which one answers a given query is a runtime decision driven by current energy state, predicted availability, and the task's objectives. The same agent logic then spans satellites with embedded GPUs and platforms with serious onboard accelerators.
A first-class layer of the stack.
This adaptive serving layer can't be an afterthought bolted onto the agent. It has to integrate with power management, task scheduling, and comms — and reason about worst-case execution and peak power, not averages. LEO today, cislunar and deep space next: the environments differ, but the constraint is the same. Reliable agents under extreme, time-varying energy limits.
NEXT INSIGHT Toward trainable AI agents for satellite operations →