DPhi runs LLM aboard satellite

Satellites are getting a lot better at collecting data. The weak link is getting that data back down fast enough to do something useful with it. That is the gap DPhi Space is trying to close, and last week it showed the clearest version yet: a vision-language model ran directly aboard its Clustergate-2 compute payload in orbit and generated a caption for an Earth image. Not on the ground. Not after download. In orbit, on the satellite itself. ### What actually ran in space? The model was Liquid AI’s LFM2-3B, a vision-language model that can look at an image and describe it in text. During commissioning, the host spacecraft passed an Earth image to Clustergate-2, which processed the image onboard and produced a multi-sentence description. DPhi framed that as the first public demonstration of on-orbit large language model inference on its platform. ### What is Clustergate-2? Basically, it is a compute server built to live in space. DPhi describes it as a programmable, cloud-like platform for satellites, with a heterogeneous processing cluster, support for containerized software, and an NVIDIA GPU supplied by EDGX. The point is not one flashy demo. The point is to let developers deploy software in orbit more like they deploy software in the cloud. ### Where is this thing flying? Clustergate-2 launched at the end of March 2026 on SpaceX’s Transporter-16 rideshare mission and is hosted aboard Momentus’ Vigoride 7 orbital vehicle. That detail matters because DPhi is not building a giant standalone satellite here. It is building shared orbital infrastructure — something other companies can plug into for software and payload experiments. ### Why bother running the model onboard? Because downlink is the bottleneck. Space sensors can gather more imagery and telemetry than satellites can realistically beam back to Earth. If a satellite can classify, summarize, or prioritize data before transmission, it does not have to send everything home raw. Think of it like moving the first pass of analysis from a crowded call center to faster decisions. ### Why is a captioning demo a useful test? Because it forces the system to do two things at once — understand imagery and turn that understanding into language. That is more interesting than a simple image classifier, which only picks from preset labels. A vision-language model is closer to the kind of flexible onboard assistant people actually want: one that can explain what it sees, from the platform design and the demo choice, but it fits the roadmap DPhi is signaling. ### What could this be used for next? DPhi and Liquid AI are already pointing developers toward applications like maritime anomaly detection and illegal mining identification. The company’s public docs also mention access to orbital telemetry, satellite health data, camera feeds, and LiDAR. So the bigger play is not “a satellite wrote a caption.” It is “developers can ship software to orbit and run AI against live spaceborne data.” ### Is this really a big shift? Potentially, yes — but the catch is scale. One successful demo does not prove that large models will become routine across satellites, where power, thermal limits, radiation tolerance, and bandwidth still bite hard. But it does show the stack is moving from theory to operation. A few years ago, “run cloud-style software in orbit” sounded like a slide deck. Now there is a live payload doing it. ### Bottom line? This is really a story about where space computing is headed. DPhi’s demo matters less because the model was huge or poetic, and more because the inference happened next to the sensor. If that keeps working, satellites stop being dumb cameras with radios and start becoming systems that can decide what matters before Earth ever sees the raw feed.