orbital inference

Orbital inference refers to satellite-mounted computers handling the task of sending queries to an already-trained AI model and receiving answers, all while orbiting Earth. In simple terms, it's like launching an entire data center building into space.

Ground-based data centers need electricity pulled from the grid, water or air cooling systems, and large plots of land. Satellites in orbit, on the other hand, can draw power from sunlight and use the cold vacuum of space to dissipate heat — that's the core idea behind this approach. But building a satellite and actually getting it into space are two very different challenges. If there's no room on a rocket, even the best-designed satellite stays grounded.

Cooling also works differently in space since there's no air. Chips carried on satellites need to be designed with everything considered together: the size of radiators needed to release heat, shielding against cosmic radiation, and durability to survive the shock of a rocket launch.

News articles use it in phrases like "plans to carry out orbital inference work for U.S. government agencies and others." It's worth noting that inference here doesn't mean training a new AI model — it refers to the stage where an already-trained model is actually run in orbit to produce answers.