A startup wants to be the Android for satellites running their own AI

When a SpaceX rocket lifts off this week carrying Google's first orbital data center prototype, it will also be carrying a smaller payload with a different theory about who gets to build AI in space.

Satlyt, a two-year-old startup that builds no satellites of its own, is flying its software on the same mission. The company says its code can run a small language model directly on a satellite, letting the spacecraft make decisions in orbit instead of downlinking raw imagery and sensor logs to ground stations for analysis. Founder Rama Afullo calls it "the Android of orbital data centers," a horizontal software layer that any operator can install regardless of who built the bus or who owns the constellation.

The bet drew an $8 million seed round led by Non Sibi Ventures, TechCrunch reported Thursday. Two prior software demonstration missions have flown, including a Gemma deployment on a Momentus spacecraft that used a quantized version of Google's Gemma 3 1B model running through llama.cpp to process logs, errors, and stack traces from onboard image-processing workloads.

The strongest reported number so far is a reduction in diagnostic payload size. In two fault-injection scenarios described in Google DeepMind's case study, Satlyt's setup compressed diagnostic output from 1,319 to 469 bytes (64.4%) and from 1,318 to 464 bytes (64.8%), with generation rates of 22.71 and 25.48 tokens per second respectively. Those are vendor-reported representative benchmarks, not demonstrated fleet-wide savings, and the case study does not establish how the compressed diagnoses compare to conventional compression or to sending the raw data down.

Afullo's pitch is that compressing and triaging data on orbit changes the economics of running constellations at all. A satellite that can summarize its own anomalies needs less downlink bandwidth, fewer ground controllers, and less round-trip latency when something goes wrong. TechCrunch reports the company is preparing software for up to 50 deployments and has set a model-memory reduction target as part of its roadmap. None of that is operating revenue.

The "Android" framing is the structural argument: orbital AI is not a closed race between SpaceX and Google because the software layer that runs on top of them is open. Afullo told TechCrunch he pitched the idea inside both companies before starting Satlyt. Both said no. This week's launch puts that thesis in the same payload bay as Google's Project Suncatcher prototype, a Google-built orbital compute experiment, not a Satlyt product, but a useful proof that the hardware side of the stack is moving.

The caveat is real. Orbital data centers are still pre-commercial. Energy generation in orbit, heat dissipation in vacuum, and launch economics all remain unresolved at any meaningful scale, and Satlyt's evidence so far is two demo missions and one benchmark. The company depends on SpaceX for launch and rides alongside a Google prototype it does not own. A planned TakeMe2Space mission will test NASA cloud protocols and Stellerian image-processing workloads, and a proposed cloud spanning two satellites is expected next year, scheduled, not shipped.

What this week's launch actually tests is whether the horizontal-layer thesis survives contact with orbit. If Satlyt's software runs cleanly alongside Google's prototype and the diagnostic savings hold up under real fault conditions, the argument that orbital AI is a platform category, not a constellation owner's private moat, gets its first live data point. If it doesn't, the seed round will look like what the skeptics expect: a demo, not a business.