The discussion of space is going from launch cadence to compute capacity. And at Disrupt 2025, government, start-ups and the tools layer leaders were in agreement: the advantage is increasingly about AI doing its thing on-the-edge — turning raw orbital EO data into insight almost in real time.
Onstage, The Aerospace Corporation CTO Dr. Debra Emmons and Ursa Space Systems CEO Adam Maher joined Violet Labs co-founder Dr. Lucy Hoag to discuss how on-orbit AI, autonomous constellations, and AI-native engineering pipelines are reinventing mission economics and speed of decision.
On-Orbit AI Transitions From Concept To Capability
Space is a bandwidth-constrained environment. Downlink bottlenecks convert terabytes’ worth of imagery into tactically delayed decisions. The patch is uploading models to the spacecraft. European missions like PhiSat-1 demonstrated onboard inference capable of discarding unusable frames, with early testing suggesting the feature could cut downlink waste by about 30% — a proof point now scaling across commercial fleets.
Emmons pointed out the architectural change: resilient, updatable compute stacks that fly along with sensors. NASA’s High-Performance Spaceflight Computing (HPSC) program, developed with industry partners, seeks 100× performance over heritage systems, allowing neural nets for cloud detection, wildfire identification and anomaly triage to be deployed where the data is captured.
Pixels to Decisions in Minutes With Onboard AI
Maher described how synthetic aperture radar (SAR) feeds convert from raw sensor data to operational intelligence when AI performs change detection, object classification and prioritization prior to downlink. Ursa’s customers are using these outputs to keep an eye on oil storage, ports and other critical infrastructure — “applications where minutes actually count,” in Maher’s words, and where weather-agnostic SAR has a distinct advantage.
Operational pilots over commercial constellations consistently demonstrate that pre-filtering and event-driven compression can reduce data volumes 70–90% without losing high-value detections. That translates into more frequent revisits, quicker incident response and reduced boots-on-the-ground costs — advantages now being tapped not just by analysts but also by emergency managers and insurers.
Independent Constellations And Space Safety
AI is also turning satellites into smarter operators. Dynamic tasking is enabled through reinforcement learning and predictive models — retargeting sensors by the dynamic path of weather elements, maritime patterns, or movement/radiation spread across a wildfire. Emmons cited prototype efforts that combine onboard analytics with crosslink coordination for clusters to share detections and reduce duplicate collects.
With the global catalog tracking in excess of 50,000 objects at any one time, machine learning is finding its way into space traffic management for triage of conjunction alerts and probability estimation.
Onboard autonomy can provide a second layer of robustness, making it possible to decide on maneuvering quickly even when communications are inhibited.
The Digital Thread That Enables Space to Construct More Quickly
AI is flying — not only that, it’s designing. Hoag explained how Violet Labs is weaving a “digital thread” through CAD, simulation, test and supply chain data to eliminate the spreadsheet glue that gums up complex hardware programs. The result: fewer integration surprises and more transparent configuration control from concept to launch.
Other sector teams transitioning to model-based systems engineering, supported by AI for requirements traceability and failure-mode analysis, have achieved double-digit decreases in rework and cycle time. For space startups racing on cadence, that could be the difference between leading a market and seeing it pass by.
Economics And Infrastructure Catch Up In Orbit
Decreasing launch costs, common smallsat buses and radiation-tolerant compute are converging with cloud-to-space pipelines.
Ground segments are increasingly able to treat satellites as edge nodes that sync with terrestrial AI stacks for retraining and version control. It’s a playbook familiar to anyone who has spread an AI capability across fleets of IoT devices — except for the orbital constraints.
Security and trust are on the agenda again. Speakers emphasized the necessity of certified models, formalisms for autonomy and secure over-the-air updates. Government standards bodies and industry consortia can be expected to advocate for test frameworks that demonstrate AI behavior under radiation upsets, sensor drift, and adversarial conditions.
What To Watch Next In On-Orbit AI And Autonomy
Edge-optimized end-to-end foundation models trained on multispectral and SAR data will advance multi-mission payloads. Federated learning between satellites might allow for models to remain fresh, without data moving. And as crosslinks multiply, constellations will serve much like distributed computers — distributing tasks to wherever power, view and bandwidth align.
The message from Disrupt 2025 was clear: AI is no longer an add-on for space — it’s the operating system. Those who combine on-orbit intelligence with AI-native engineering and secure update pipelines will dominate the decade ahead.
