Satlyt's $8M Bet: Running AI Models on Satellites Already in Orbit

Satlyt has raised an $8 million seed round to run AI models on satellites in orbit. The round was led by Houston-based Non Sibi Ventures, whose partner Bernard Harris served as a NASA astronaut for more than 20 years. The financing was disclosed on Oct. 1, 2026. TechCrunch

The company was co-founded by Rama Afullo after he left SpaceX. Afullo, a Kenyan American engineer, previously worked in Google's cloud computing business before a brief stint at SpaceX's Starlink in 2024. Satlyt maintains headquarters in Sunnyvale, California and Nairobi.

Satlyt does not build spacecraft. It builds software to operate AI models on satellites owned by others. Its stated aim is to let many satellites share the work of large computing jobs in space. It does not sell buses or payloads, the standard terms for satellite bodies and the equipment they carry.

The company has flown its software on two demonstration missions. Earlier this year, it deployed Google DeepMind's Gemma model, a lightweight AI model built to run efficiently, aboard a spacecraft operated by Momentus. In that flight, onboard processing cut the size of a transmission about onboard software errors by more than 60%. TechCrunch

Its next software flight is scheduled to launch on a SpaceX rocket alongside the first prototype for Google's Project Suncatcher space data center effort. The host spacecraft was built by TakeMe2Space, an Indian startup that builds computing hardware for satellites. That flight is a multi-tenant technology mission, meaning several groups share one satellite to test different ideas. NASA will use it to test protocols for cloud computing in space, Stellerian will test image processing for space surveillance, and TakeMe2Space will test hosting other companies' software.

Satlyt expects to attempt next year to create a shared computing system spanning two different satellites. The plan is disclosed as an attempt, not a completed capability.

The broader context here is familiar from cloud work on the ground. The downlink, the radio link from satellite to Earth, remains the choke point for Earth observation and in-orbit telemetry, the routine health data satellites send home. Raw sensor output and logs are expensive to move, store and sort on the ground. On-orbit inference, running AI where data is collected, inverts that workflow to filter at the edge and transmit the insight. Limits include tight power, heat limits, radiation-induced bit flips, mixed processors, and patchy links.

In my view, the Momentus result is the most instructive detail disclosed so far. A 60% cut in transmitted bytes for a diagnostic task is not autonomy or full onboard analytics. It is log compression through semantic filtering. That is a practical entry point because operators understand bandwidth savings and can price them. More complex jobs across satellites can follow once flight heritage, tooling and trust accumulate, which would let operators get more from data they already collect.