Satellites Are Now Reading Earth in Orbit — Loft Orbital Picked by NASA JPL to Run 'Gemma 3 in Space'

From "snap-and-downlink" satellites to "see-and-decide" satellites

Here's the deal: on June 23, Loft Orbital was selected by NASA's Jet Propulsion Laboratory (JPL) to validate Earth-science AI software — in space. This isn't about launching one more satellite. It opens a real proving ground for "space AI": satellites running AI in orbit to interpret Earth themselves.

There's an already-proven punch behind it. In April 2026, Loft's YAM-9 satellite ran Google DeepMind's Gemma 3 vision-language model (VLM) directly in orbit. Running the model on an onboard processor (Nvidia Jetson Orin AGX class), the satellite answered natural-language queries like "where the natural environment meets human development" or "infrastructure around railway hubs," classifying imagery itself. For the first time, a satellite judged "what it saw" without downlinking raw pixels to the ground.

Why is this big? Until now, Earth-observation satellites were cameras. They downlinked vast raw datasets, and only then did ground analysts and computers look. But when a satellite decides in orbit, data latency disappears entirely. Disasters like wildfires and floods shift from "analyzed days later" to "detected in real time." JPL choosing to validate this with its own AI software means it's taking the paradigm seriously.

So today's story: what Loft Orbital does and what YAM-9 achieved, what JPL's "NAVI-Orbital" software is, what changes when a satellite decides for itself, and what this throws into the space and AI industries. Start with the cast.

The cast — Loft Orbital, NASA JPL, and Gemma 3 in orbit

First, Loft Orbital. A company that rents out "satellite infrastructure as a service." It operates satellites and lets customers "load" their own sensors or software to run in space — think "cloud hosting in orbit." Customers don't build a satellite from scratch; they just put their mission module on Loft's satellite. YAM-9 is one such platform satellite.

Next, NASA JPL (Jet Propulsion Laboratory). NASA's flagship lab, behind Mars rovers and deep-space missions. What JPL brought here is software called "NAVI-Orbital" — a zero-shot vision-language-model system that makes a satellite answer natural-language queries directly. JPL wanted to validate it in the real space environment and chose Loft's infrastructure as the testbed.

Third, Gemma 3 in orbit. Google DeepMind's lightweight open model. The key is it's light enough to run on a small onboard chip on a satellite, not a giant cloud GPU. Not a heavy, expensive frontier model — a "light enough to run anywhere" model proving its worth in the extreme environment of space. The ultimate edge AI is, it turns out, orbit.

Tie the three together: on a satellite from a company that rents space infrastructure (Loft), a lightweight AI model (Gemma 3) runs directly to validate NASA's Earth-science software (JPL's NAVI-Orbital) in space. That's the spine.

What was actually announced

Words scatter, so here are the confirmed facts.

Item	Detail
Selection	June 23, 2026 — NASA JPL picks Loft Orbital
What's validated	JPL's Earth-science AI software "NAVI-Orbital"
Platform satellite	Loft Orbital YAM-9
Prior milestone	April 2026, ran Gemma 3 VLM in orbit on YAM-9
Onboard hardware	Nvidia Jetson Orin AGX-class processor
Core capability	Satellite classifies imagery via natural-language query (no raw-pixel downlink)
Flight demos	Begin June 2026; further deployments 2027–2028
Target use	Real-time detection of wildfires, floods; removing data latency

Line by line. First, the "no raw-pixel downlink" point is decisive. Legacy satellites had to transmit all captured data to the ground, and limited bandwidth was always the bottleneck. When the satellite decides "this scene is a flood" in orbit, it downlinks the conclusion instead of the vast raw data — saving bandwidth and time at once. A flip in framing.

Second, the natural-language interface is the innovation. Classifying satellite imagery used to require training a dedicated model per object. NAVI-Orbital, being VLM-based, understands worded queries like "find infrastructure around railway hubs" on the spot — finding even un-trained objects zero-shot. The satellite gained not just "eyes" but "an ear for language."

Third, the phased roadmap. Flight demos begin June 2026, with further deployments on AI-enabled satellites in 2027–2028. To cover anywhere on Earth in real time would take 50–100 satellites like YAM-9, by Loft's reckoning. Beyond a single successful experiment lies a long-term vision of growing into a constellation.

What each side gets

Loft Orbital's win first. One, it landed NASA JPL — the most authoritative reference. "To validate space AI, use Loft's infrastructure" becomes a powerful calling card for other government and enterprise customers. Two, it's the decisive proof of its "satellite as a service" model — a customer running its own software (NAVI-Orbital) on Loft's satellite means the promised "orbital hosting" actually works.

NASA JPL's win is clear too. It gained a fast path to validate its AI software in the real space environment without building a satellite from scratch. For missions where speed is life — disaster response — onboard AI that removes data latency is a game-changer. If you can catch a wildfire the instant it starts rather than days later, the value is hard to price.

And the surprise beneficiary is Google DeepMind. Its lightweight Gemma 3 "running even in space" is an extreme edge-AI case. Proof that it works in the farthest place the cloud can't reach is powerful marketing that Gemma-class open, lightweight models reach beyond terrestrial IoT and robots into orbit. On a different axis from the heavy frontier-model race, it proved "the value of being light."

Past parallels — wins and failures

Edge AI reshaping the board through "on-the-spot decisions" already has terrestrial precedents. The closest is self-driving cars. A car can't wait to ask the cloud, so it judges what its cameras see, in the car, instantly. Onboard satellite AI follows the same logic — where comms latency is fatal, "decide on the spot" is the only answer.

The textbook success is shrinking data to grow value. Instead of sending all the raw data, send only what's meaningful from the field, and bandwidth, cost, and time all drop. A satellite downlinking only "flood detected" is the archetype; if it works, the economics of satellite operations change wholesale.

Conversely, the risk of failure is real. Space is extreme — radiation, power, heat — so a model that runs well on the ground isn't guaranteed to operate stably in orbit. And when a satellite's self-made judgment is wrong, verifying and correcting it is tricky. If YAM-9's April experiment showed "it can be done," this JPL validation is the real test of "how much can it be trusted."

Competitor counter-plays

The counter from traditional large satellite/aerospace firms is "reliability and scale." Already operating many observation satellites, they can follow with proven stability and vast data assets: "we can put AI up too." But Loft's agility — quickly and flexibly loading customer software like "satellite as a service" — is a differentiator that's hard to imitate.

Other lightweight-model camps, watching this, will eye the space and edge markets. Gemma 3 took the first in-orbit run, but Llama-family or other open lightweight models could target the same seat. The title "the model that runs in the most extreme environment" is a powerful flag in the edge-AI race.

And cloud providers sit in an odd spot. Satellites deciding for themselves may shrink some demand for "downlink everything and analyze in the cloud." But conversely, "ground backend" demand to aggregate satellites' conclusions and run bigger analyses still belongs to the cloud. The natural picture is a hybrid where "field judgment (satellite)" and "deep analysis (cloud)" split roles.

So what actually changes

If you're in disaster response or the public sector, this is direct good news. If satellites detect wildfires, floods, and earthquakes in real time, response time could drop from days to minutes — the potential to turn "waiting for data and missing the golden hour" into "alert the instant it happens." Until reliability validation is complete, though, treat it as a supplementary tool.

If you're a space or AI startup, note that "the barrier to entry is dropping." Without building a satellite from scratch, you can validate your AI software in space just by loading it onto a platform like Loft's. The formula "space = only for giant capital" breaks, and opportunity opens to small, software-centric teams.

If you're a general reader, the meaning is "AI is leaving the cloud for the field." The AI we use usually lives in giant data centers, but real innovation often happens when "AI decides for itself in the hardest place to reach." A satellite reading Earth from orbit is the most dramatic example of that direction.

One more layer — the "economics of bandwidth" and space edge AI

The hidden core here is the "economics of bandwidth." An Earth-observation satellite's biggest cost and constraint isn't the camera — it's the downlink. The data a satellite captures is enormous, but the comms bandwidth to send it from orbit to a ground station is limited and expensive. So satellites have always suffered the bottleneck of "barely downlinking a fraction of what they can capture." When a satellite decides in orbit and sends only "the conclusion," that bottleneck disappears entirely — a fundamental redesign of satellite economics, extracting far more value from the same comms budget.

The meaning of a lightweight model like Gemma 3 comes back into view here. Space is the extreme of a triple constraint: power, heat, radiation. A GPU drawing hundreds of watts like a data center can't fly; the model must run on a small onboard chip. So not "the smartest model" but "a light model that's smart enough within the constraints" becomes the answer. If the frontier-model race is a fight over "size and performance," space is the opposite — a fight over "how small you can shrink it and still be useful" — and Gemma 3 just posted the first victory.

And it's worth chewing on that NASA JPL chose to validate by "renting Loft's satellite." In the old days, JPL designing, building, and launching its own satellite alone took years and a huge budget. But loading just its software onto a "satellite as a service" shrinks the validation cycle from years to months. This changes the very speed of space R&D — a signal of moving from owning hardware to "subscribing" to space infrastructure.

But beneath all this rosy picture lies the heavy homework of "reliability." If a satellite ruled "this is a flood" and was wrong, the crux is how the ground catches and corrects that misjudgment. Not sending raw pixels is the key to efficiency — yet that creates a paradox: there's no raw data on the ground to re-verify against. So the real question of this JPL validation isn't "does it work" but "how much can you trust and delegate," and "how do you reverse a wrong call." The future of space edge AI will be decided not by performance but by the design of this trust.

🥄 Three Things You're Probably Wondering

— If a satellite decides for itself, can't it be wrong? Of course. Space is extreme — radiation, power limits — so even a ground-proven model can wobble in orbit. That's why the crux of this JPL validation isn't "does it work" but "how trustworthy is it." Reliability is still under test, so it's too early to call.

— What does this mean for me? Nothing direct right now. But if disaster alerts for wildfires and floods get faster, that ultimately ties to our safety. Smarter satellites loop back into ground safety systems.

— Why a small model like Gemma 3 and not a giant one? The chip on a satellite is incomparably smaller than a data-center GPU, with tight power. So a "light enough to run anywhere" model is the answer. In space, "does it run" comes before "is it smart."

Further reading

• Loft Orbital Selected by NASA's Jet Propulsion Laboratory to Deploy AI Software for Earth Science Applications — BusinessWire
• A satellite just learned to find things on its own — here's what that means — TechCrunch
• Satellite AI Inference Clears Orbit: Gemma 3 Ran Aboard YAM-9 in April — TechTimes
• Loft Orbital's YAM-9 Runs Google DeepMind's Gemma 3 Directly in Space — AlphaSignal
• NAVI-Orbital: First In-Orbit Demonstration of a Zero-Shot Vision-Language Model for Autonomous Earth Observation — arXiv

Numbers and criteria are as of announcement and may change.