Under a June 23 agreement, Loft Orbital and NASA's Jet Propulsion Laboratory begin testing whether a satellite can use onboard AI to cue another satellite — instead of beaming raw imagery home first.
The bottleneck that kept Earth-observation satellites as passive raw-data relays has just broken. Small, open-source multimodal AI models, the kind that fit on a laptop, can now run inside the cramped, radiation-hardened computers bolted to a satellite bus, and a NASA-funded project called Federated Autonomous Measurement (FAME) is among the first government programs designed to test that capability on orbit.
On June 23, Loft Orbital, a San Francisco-based "satellite-as-a-service" operator that flies customer payloads, announced an agreement with NASA's Jet Propulsion Laboratory to run JPL's AI software on a Loft Orbital spacecraft under the FAME program, which is funded through NASA's Earth Science and Technology Office (Loft Orbital announcement via SpaceNews). Tests started this month, with follow-on runs planned for 2027 and 2028 on additional Loft Orbital spacecraft.
What FAME is trying to prove is a small operational change with large consequences. Instead of treating satellites as cameras that dump raw imagery to ground stations, FAME wants one satellite to look at what it has already seen, decide what is interesting, and tip off another satellite to image the same place in more detail or with a different sensor. Today that "tip-and-cue" workflow almost always requires the imagery to be downlinked to the ground first, where human analysts or ground-side models pick the next target. Under FAME, the model that picks the next target runs onboard.
The economic driver is bandwidth. Modern Earth-observation constellations generate far more pixels than they can economically beam back to the ground, so most of what they see is either discarded, summarized coarsely, or never tasked in the first place. If a satellite can throw away ninety percent of its images and keep the ones a downstream spacecraft should also look at, the data that does make it to the ground is much more valuable per byte. "The goal is to capture, sense, understand, and send insights about Earth without downlinking big amounts of data," said Paul Lasserre, general manager for AI at Loft Orbital, in an interview (SpaceNews).
The reason this is a 2026 milestone rather than a 2016 one is that the models themselves have finally gotten small enough. Multimodal models trained on images, text, and sensor data have been getting cheaper to run for years, but Earth-observation satellites operate under tight compute constraints: low-power processors, limited memory, and hardware that has to survive radiation and thermal cycling without crashing. Lasserre said the AI models are trained on a "very large corpus" of data and that the size-versus-capability curve has only recently bent enough to make on-orbit inference practical (SpaceNews). His framing is deliberately hedged: the compute constraint on satellites is real, he said, "but it's not really limiting." That is the source's own hedge and should be treated as such, not paraphrased into certainty.
FAME is a funded experiment, not a deployed capability. There is no public Earth-science result yet: no flood map, no wildfire detection, no ice-sheet measurement has come out of the on-orbit inference path, only the promise that the architecture works at all. Independent claims about accuracy gains, latency improvements, or "first of its kind" status are not supported by the public source set and should be treated as forward-looking until FAME produces results.
There are also second-order questions the FAME test does not yet answer. The onboard models that would pick which imagery matters are typically small, open-source multimodal models, which means their behavior is shaped by the broader open-source AI community rather than by NASA or Loft Orbital alone. Auditing why a model flagged a volcanic plume as worth a closer look, or missed a methane leak entirely, is a different problem from auditing a ground-side analyst's queue. And running inference reliably under space radiation, thermal swings, and a multi-year mission lifetime is an engineering question the FAME team has not publicly answered; the first round of tests is meant partly to find out where the model breaks before more capable follow-on missions depend on it.
The institutional shape of FAME also matters. It is funded through NASA's Earth Science and Technology Office (NASA ESTO), the part of the agency's Earth-science portfolio that pays for the technology behind future missions rather than flying the missions themselves. ESTO's bet is that on-orbit inference becomes a standard building block for the next generation of Earth-science constellations, the way onboard GPS receivers or solid-state data recorders once did. If FAME works, the next decade of NASA Earth-science solicitations is likely to assume that tip-and-cue is solved at the spacecraft level, not at the ground segment.
Loft Orbital's role is also part of why this announcement reads as a contract story but functions as a category moment. Loft Orbital sells hosted-payload space, with customers buying a slot on one of its buses instead of buying their own satellite, and that model depends on the company being willing to fly experiments like FAME alongside commercial payloads (Loft Orbital). For NASA, that means getting an on-orbit test bed without flying a dedicated mission. For Loft Orbital, it means its buses become the reference platform for the kind of compute-heavy Earth observation the rest of the industry will be watching.
The watch items through 2027 and 2028 are narrow and concrete. First, does the FAME test produce any public Earth-science result, a published map, a benchmark, or a real tip-and-cue event captured end-to-end, or does it remain a flight-qualification exercise? Second, do the follow-on Loft Orbital missions expand the test from a single satellite pair to a small federated constellation, which is the regime FAME is named for but has not yet operated in? Third, does the small open-source model the system leans on stay current with the open-source community, or does its accuracy drift as the broader model ecosystem moves on? None of those answers are in the public record yet. The June 23 announcement is the test going live, not the result.