For as long as Earth-observation satellites have photographed the planet, they have operated the same way: capture imagery, beam the raw files to a ground station, and wait for analysts or machine learning pipelines running in terrestrial data centers to extract meaning. The satellite is a camera bolted to a radio transmitter — sophisticated hardware, but fundamentally passive.

In April, that architectural assumption broke. Loft Orbital, a Franco-American space infrastructure company, ran what TechCrunch reported June 15 as the first publicly disclosed execution of a vision-language model in orbit. Aboard Loft’s YAM-9 spacecraft, Google DeepMind’s Gemma 3 answered natural language queries, classified Earth imagery, and produced plain-English summaries of what it found — without transmitting a single raw pixel to the ground first. The satellite had stopped being a passive sensor and had started being an analyst.

That shift is what makes this demonstration significant: the value layer in Earth observation has moved. It is no longer primarily located in the sensor, or in the downlink, or in the ground-based processing pipeline. It is now located in the inference capability running at the satellite itself.

What YAM-9 Actually Did in April

YAM-9 launched November 28, 2025, aboard a SpaceX Transporter-15 rideshare mission as Loft’s pathfinder for orbital AI. The spacecraft carries an Nvidia Jetson Orin AGX graphics processing unit — one of the highest-performance edge accelerators currently flying in low Earth orbit.

The model that ran on it, Gemma 3, is an open-weight, decoder-only transformer from Google DeepMind, available in sizes from 1 billion to 27 billion parameters. Its architecture uses a SigLIP vision encoder that converts imagery into token embeddings, which the language model then processes alongside text — allowing it to receive a question in plain English about an image and return a natural-language answer. The 5:1 ratio of local-to-global attention layers in Gemma 3’s design reduces KV-cache memory pressure at long context windows, a property that matters significantly for memory-constrained orbital hardware.

But Gemma 3 alone does not run in space. The critical engineering contribution was NAVI-Orbital, a software harness developed by Juan Delfa Victoria and the artificial intelligence group at NASA’s Jet Propulsion Laboratory. Gemma 3, even in its smaller configurations, carries more software library dependencies than a radiation-hardened space computer can support without modification. NAVI-Orbital stripped those dependencies down and restructured the inference pipeline to fit within YAM-9’s memory budget and processing constraints. LangGraph, an agent orchestration framework, managed the autonomous workflow onboard — sequencing the queries, passing imagery to the model, and routing the model’s outputs.

The result: researchers on the ground issued natural language commands and received structured answers without a real-time data downlink. Queries included classifying areas where natural environments meet human development and identifying infrastructure around railway hubs. The satellite interpreted its own sensor data, flagged what was relevant, and returned summaries in plain English — from 500 kilometers up.

The project originated, according to Delfa Victoria, with a JPL researcher named Taran Cyriac John who was thinking about voice-interactive assistants for pressurized-suit astronauts on the Moon or Mars. “How about we provide an assistant, like in video games and in movies, where you see an AI which is interactive?” Delfa Victoria told TechCrunch. YAM-9’s April demonstration is the first operational expression of that idea.

Why the Satellite Downlink Bottleneck Matters

To understand what YAM-9 changed, it helps to understand what existed before. A single high-resolution Earth-observation satellite produces 1 to 2 terabytes of imagery per day. Its contact window with any given ground station lasts 5 to 15 minutes per orbital pass. The physics of low Earth orbit mean a satellite can contact a ground station in only four to six windows per day — each just ten minutes long from horizon to horizon. At radio-frequency downlink speeds, only a fraction of the day’s data reaches the ground before new data overwrites it.

The structural inefficiency compounds. Research has consistently found that roughly 67 percent of raw satellite imagery is obscured by clouds and therefore low-value — yet under the conventional bent-pipe architecture, all of it gets downlinked regardless. Analysts on the ground then sort through the entire volume to find the useful fraction.

An on-orbit VLM that can run initial triage — discarding clouded frames, classifying relevant scenes, flagging anomalies — before downlinking compresses that problem fundamentally. Instead of terabytes of raw imagery, the satellite transmits megabytes of structured findings and flagged crops. The ground station receives intelligence, not raw data.

The economic implications for disaster response are direct. When a wildfire ignites or a river delta floods, the hours between a satellite pass and an actionable ground-station report determine whether responders have current information or are operating blind. An orbital reasoning system that can autonomously identify affected areas and transmit only its analysis eliminates that latency window.

What Owning the Orbital Inference Layer Means

Paul Lasserre, Loft Orbital’s head of AI, described the defense and ISR implications directly to TechCrunch: “If you have a VLM, you can have logic — like ‘monitor this border for me, and let me know when something is suspicious,’ and interact back and forth with the satellites.” That is a qualitative change in how intelligence, surveillance, and reconnaissance missions operate. Persistent, conversational tasking of an orbital asset — without human analysts reviewing each frame — represents a new operational tier between raw sensor and finished intelligence product.

Loft has already translated this capability into commercial partnerships. The company has a strategic agreement with Helsing, the European defense AI company, to deploy a multi-sensor satellite constellation delivering real-time situational awareness for governmental and military customers. Hardware for that constellation was already in production as of early 2026.

For non-defense customers, the analogy is structural: a satellite that can autonomously classify and prioritize its own imagery is worth more per orbit than one that generates undifferentiated raw data. Agricultural monitoring firms, environmental watchdogs, and maritime domain awareness operators all benefit from faster triage and smaller downlink volumes. Companies that deploy orbital inference capability own the layer that controls what their customers can ask of their satellites — a new kind of competitive moat in the Earth-observation market.

Planet Labs, which operates a constellation of several hundred Dove and SuperDove CubeSats, already flies satellites equipped with the same Nvidia Jetson Orin processors as YAM-9. A Planet Labs spokesperson confirmed to TechCrunch that the company is currently using those processors for simpler object detection tasks and that research is underway on more advanced AI applications, including VLMs. Kiruthika Devaraj, Planet Labs’ vice president of engineering, described the strategic direction in a separate interview with IEEE Spectrum: “We want to put the brain, all the compute, right next to the sensors, so that the system of satellites we build acts like a biological network that is responding to stimuli in real time.”

What Orbital Edge AI Still Cannot Do

YAM-9 is a pathfinder, not a production system. Loft Orbital’s head of AI told TechCrunch that achieving real-time coverage of anywhere on Earth would require somewhere between 50 and 100 satellites of comparable capability — a constellation that does not exist.

The engineering constraints that YAM-9 navigated are also the constraints that every future orbital AI deployment will have to solve at larger scale. Power budgets in space are tight: adding a high-performance GPU inflates both mass and electrical demand, and the Jetson Orin AGX is already near the upper end of what a small satellite bus can support without significant redesign. Thermal management is a related constraint — orbital hardware cannot shed heat through convection, only through radiation.

Software lifecycle management in orbit is fundamentally different from software lifecycle management in a data center. Satellite firmware updates require narrow downlink windows and careful version control across a fleet that cannot be physically accessed for maintenance. As better versions of Gemma 3 or successor models emerge, updating them to a constellation of 50 to 100 spacecraft is a non-trivial operational challenge.

And the most significant outstanding gap: no independent performance data has been publicly disclosed for YAM-9’s VLM runs. Latency figures, classification accuracy, and bandwidth savings are not in the public record. The demonstration establishes that the architecture works; it does not yet establish how well it works at the task level, or how it degrades under cloud cover, sensor noise, or model drift over time in orbit.

A Race Now Officially Underway

The April demonstration has made the competitive landscape explicit. Kepler Communications, a Canadian company, commissioned what it describes as the world’s largest orbital compute cluster in March 2026: 40 Nvidia Jetson Orin modules distributed across 10 satellites and interconnected through laser optical links. Kepler declined to tell TechCrunch whether it had deployed VLMs in orbit due to NDA agreements with partners, but acknowledged “several undisclosed use cases” of its compute environment since those spacecraft launched in January.

The lessons learned deploying smaller models in orbit will directly inform how the industry attempts to deploy larger-scale compute infrastructure in space — particularly in the areas of power and memory management that Kepler, Loft, and incoming competitors such as Vast Space are all navigating at once.

For the space industry, YAM-9’s April demonstration occupies roughly the position that ESA’s Φ-Sat-1 mission occupied in 2020, when that spacecraft became the first to run a convolutional neural network in orbit for cloud detection — a proof of concept that trained the field on what was possible. Vision-language models represent the next capability tier: not just filtering out low-value frames, but understanding what is in them and answering questions about them in natural language.

The satellite has stopped being a camera. The question now is how quickly it becomes a full analyst.

Frequently Asked Questions

What is a vision-language model, and why does running one in orbit matter?

A vision-language model is an AI system that jointly processes images and text, allowing it to analyze imagery and respond to natural-language questions about what it sees. Running one in orbit matters because it eliminates the need to downlink raw imagery before extracting intelligence from it. Instead of shipping terabytes of unprocessed data to the ground, an on-orbit VLM can classify scenes, flag anomalies, and return structured summaries directly — compressing the gap between observation and actionable insight from hours to seconds.

What specific architecture made the YAM-9 demonstration possible?

Three components worked together: Google DeepMind’s Gemma 3 (an open-weight VLM with a SigLIP vision encoder and a 5:1 local-to-global attention ratio that reduces memory pressure); NASA JPL’s NAVI-Orbital software harness (which stripped Gemma 3’s library dependencies down to fit within the satellite’s onboard memory budget); and LangGraph (an agent orchestration framework that managed the autonomous query-and-response workflow). The compute hardware was an Nvidia Jetson Orin AGX GPU aboard Loft Orbital’s YAM-9 satellite, launched in November 2025.

How many satellites would be needed to provide real-time global coverage with this capability?

Loft Orbital’s head of AI estimates 50 to 100 satellites of comparable capability would be needed for global real-time coverage. The primary scaling challenges are power budget constraints, software update logistics in orbit, and the absence of publicly disclosed performance benchmarks from the April demonstration.

Can satellites now make autonomous decisions without human oversight?

At the level demonstrated on YAM-9, the satellite classifies and summarizes imagery autonomously in response to pre-set natural language queries — it does not take physical action. The system still depends on humans to define the queries and review the outputs. Loft’s head of AI described the near-term model as conversational tasking: “monitor this border for me, and let me know when something is suspicious.” More advanced autonomous decision-making — including any system with lethal implications — remains a governance and regulatory question that the industry has not yet answered.