The largest compute cluster in orbit is now open for business, and it’s smaller than what you’d find in a single server rack on Earth. Kepler Communications, a Canadian company, launched about 40 Nvidia Orin edge processors across 10 satellites in January, all connected by laser links, TechCrunch AI reports. The company has 18 customers and just signed its newest one: Sophia Space, a startup that will test its orbital computing software on Kepler’s constellation.
This matters because it marks the first real commercial traction for a concept that’s attracted billions in hype but precious little hardware in actual orbit.
Edge Processing First, Data Centers Later
Experts don’t expect large-scale orbital data centers, the kind SpaceX and Blue Origin have floated, until the 2030s. What’s happening now is more modest but arguably more practical: processing data where it’s collected rather than beaming everything back to Earth.
Kepler CEO Mina Mitry frames the company not as a data center operator but as infrastructure for space-based applications. “Because we have the belief it’s more inference than training, we want more distributed GPUs that do inference, rather than one superpower GPU that has the training workload capacity,” Mitry told TechCrunch AI. His key point: Kepler’s GPUs run at 100% utilization, unlike a hypothetical orbital supercomputer that might sit idle most of the time.
That distinction is worth paying attention to. It’s the difference between building a hyperscale facility in orbit (enormously hard) and running lightweight AI inference on sensor data as it’s generated (hard, but doable now).
Two Competing Visions Are Emerging
The orbital compute space is splitting into two camps:
- Edge-first players like Kepler and Sophia, focused on distributed processing, inference workloads, and serving existing satellite operators
- Hyperscale dreamers like Starcloud, Aetherflux, and the ambitions of SpaceX and Blue Origin, raising significant capital for full data center-class hardware in orbit
Sophia’s contribution could prove critical to the hyperscale camp eventually succeeding. The startup is developing passively cooled space computers, tackling one of the hardest engineering problems for orbital data centers: keeping processors cool without heavy, expensive active-cooling systems. Under the new partnership, Sophia will upload its operating system to Kepler’s satellites and attempt to configure it across six GPUs on two spacecraft. According to TechCrunch AI, this is the first time that kind of multi-node orchestration has been attempted in orbit.
The Military Connection
The U.S. military is a key early customer for orbital edge computing. The Pentagon is developing a satellite-based missile defense system that needs to detect and track threats in real time, exactly the kind of workload that benefits from processing data in orbit rather than sending it to the ground first. Kepler has already demonstrated a space-to-air laser link for the U.S. government.
This is the pattern we’ve seen before: military demand validates the technology, creates revenue, and eventually the commercial market catches up.
What’s Actually Changing
Sophia CEO Rob DeMillo made an interesting observation: a Wisconsin city just banned new data center construction, and some members of Congress are pushing similar restrictions. Every constraint on terrestrial data centers makes the orbital alternative slightly more attractive.
That’s a stretch for now. Nobody is replacing an AWS region with satellites anytime soon. But the regulatory headwinds facing ground-based data centers (power consumption, water usage, land use) are real and growing. Orbital compute doesn’t need to replace Earth-based infrastructure to be valuable. It just needs to handle workloads that are better served from space.
For AI practitioners and space-tech investors, the practical takeaway is clear: orbital compute in 2026 is an edge processing story, not a cloud computing story. Companies building for inference on sensor data have a path to revenue now. Anyone planning orbital training clusters is still years away from feasibility.
The full details are available in the original report from TechCrunch AI.