Satlyt raises $8 million and prepares to launch AI software on SpaceX Falcon 9
The startup Satlyt announced an $8 million seed round today and a planned launch this Thursday aboard a SpaceX Falcon 9. The payload will also carry the first prototype of Project Suncatcher, Google's attempt to build an orbital data center. Non Sibi Ventures of Houston led the round; its partner Bernard Harris, a former NASA astronaut with more than 20 years of service, served as the technical anchor for the investment decision.
From rejection at SpaceX and Google to an independent product
Rama Afullo, a Kenyan-American engineer who managed cloud products at Google before moving to SpaceX's Starlink division in 2024, tried to convince both employers to back the idea of orbital compute centers. Both said no. Afullo left, founded Satlyt with dual headquarters in Sunnyvale, California, and Nairobi, and chose a software-only model — no satellites of its own. His analogy: if SpaceX and Google are building the "iPhone" of orbital data centers, Satlyt will supply the "Android," an open horizontal platform that runs on many manufacturers' hardware.
An abstraction layer like VMware or Snowflake
Satlyt's software is designed to run AI models directly on the satellite, sparing the operator from managing the underlying infrastructure. The approach mirrors what VMware and Snowflake did for terrestrial cloud: abstract away the lower-layer complexity. The company has already completed two demonstration missions running code on active satellites. The initial focus is operational efficiency — cutting the need to downlink data, which is expensive, slow, and bandwidth-constrained.
Google DeepMind's Gemma cut 60% of transmission volume
Earlier this year Satlyt uploaded Google DeepMind's Gemma model onto a Momentus satellite. The model compressed onboard software error reports by more than 60%, a saving Afullo says translates to hundreds of thousands of dollars (roughly 1.1–1.5 million shekels) per satellite per year. The figure comes from a single real-world experiment, not a synthetic benchmark, and the company has not published full performance metrics for the inference run itself.
Three customers on one Indian satellite
The upcoming launch will fly on a satellite from TakeMe2Space, an Indian startup building space compute hardware. Three paying customers are on board: NASA, testing protocols for space-based cloud; Stellerian, which tracks orbital objects and wants to run local image processing; and TakeMe2Space itself, demonstrating third-party software hosting. With very few high-power GPUs currently in orbit, the immediate opportunity lies in lighter workloads like those Satlyt is demonstrating.
Next target: a cloud spanning two satellites
The company's next project, planned for next year, is to prove a shared compute system — a genuine cloud — distributed across two separate satellites. If it works, Satlyt can sell compute as a managed service to satellite manufacturers, turning every platform into a revenue generator. Afullo puts it plainly: "We turn your satellite into a managed service that generates revenue."
The investors' bet: pragmatism over maximalist vision
Non Sibi Ventures explained that the investment rests on Satlyt not depending on the most aggressive vision of orbital data centers — one that demands extreme GPU density and suffers from unclear economics and launch bottlenecks. Instead, the company is building a software layer that creates value today on existing hardware, while keeping the upside for the future.