Kepler Communications opens the largest orbital cluster yet: 40 Nvidia Orin chips

Kepler Communications has brought to market the largest orbital computing cluster to date and has already started selling access to it. The first notable customer is Sophia Space: the startup will test its space-computer software on Kepler's satellites.

What Kepler launched

Kepler's cluster consists of about 40 Nvidia Orin chips aboard 10 operational satellites connected by laser links. The company put the system into orbit in January and had signed up 18 customers by April.

This is not yet a full orbital data center. It is narrower, more practical infrastructure: relaying data between spacecraft, running computation close to the data source, and speeding up processing without constantly sending the entire data set down to Earth.

Kepler stresses that it is not building an orbital data center in the usual terrestrial sense, but a network layer for applications in space and in the near atmosphere. The company wants to serve not only its own spacecraft but also third-party satellites and, eventually, drones and aircraft that need fast connectivity and data processing without a long route through ground infrastructure.

In other words, the value lies in constant availability of compute in orbit, not in record performance from any single spacecraft.

A test for Sophia Space

For Sophia Space, the partnership with Kepler is a chance to validate a key element of its future platform before launching its own satellite, planned for late 2027. The startup is developing space computers with passive cooling. The idea is to sidestep one of the main problems of orbital computing: powerful processors run hot, and active cooling systems make a spacecraft heavier, more expensive and harder to launch.

For a project like this, testing in the real orbital environment matters more than lab tests.

According to the plan, Sophia will upload its own operating system to a Kepler satellite and try to deploy and configure it across six chips on two spacecraft at once. On terrestrial servers this task looks basic, but in orbit it will be the first experiment of its kind.

If everything works, the company will remove an important technical risk before its first launch and get practical confirmation that its architecture is suited not to demonstrations but to real operation.

"We need distributed accelerators for inference, not one

ultra-powerful chip for training."

Where demand will come from

The clearest scenario for orbital computing is not training large models in space but processing data right where it is generated. This matters especially for satellite sensors, where response time and link bandwidth are limited. If some tasks run in orbit, the satellite can send down the processed result instead of the raw data stream. That cuts latency and saves expensive downlink capacity.

• Onboard preprocessing of imagery and telemetry
• Work with "heavier" sensors such as synthetic aperture radar
• Network and compute services for third-party satellites
• Laser data links between space, aircraft and drones
• Fast event recognition and filtering for government and defense systems

Kepler has already demonstrated laser communication between a satellite and an airborne platform in a demonstration for the US government. That helps explain why the market is taking shape now: the large orbital data centers that SpaceX, Blue Origin and several well-funded startups talk about still look like a story for the 2030s. Distributed computing for specific tasks, by contrast, can be sold today, especially when customers care more about continuous workload and useful output than peak performance.

What this means

The space computing market is shifting from big promises to the first useful infrastructure. Kepler and Sophia Space are betting not on giant orbital farms but on a practical compute layer next to satellite data. If such tests become routine, space-tech will gain a new working segment well before full-scale data centers in orbit.

For the AI market, it is also a new class of infrastructure built for inference at the network edge.