Orbital AI is an infrastructure race, not a model race

Compute is moving to orbit, and the deployment gap, not the silicon gap, decides who owns the next infrastructure layer. The common read is that China is playing catch-up. The deployed hardware says otherwise: the on-orbit layer is already contested, and the race is no longer about who has the best chip but about who has the most operational satellites stitching themselves into a network.

The mechanism is on-orbit edge AI. Shift inference to where the data lands, and remove the ground-station bottleneck by tying satellites together with optical links. TechTimes' reporting on a 12-satellite Chinese constellation quantifies what "shipped" looks like: an 8-billion-parameter model running 744 trillion operations per second on commercial off-the-shelf chips inside a 6U CubeSat, comparable in scale to widely deployed commercial LLMs in a satellite roughly the size of a shoebox. The laser inter-satellite links held 99.99% data-availability uptime over roughly eight continuous days at distances up to 1,000 km, the kind of sustained link performance that lets the constellation answer over images instead of waiting to dump raw data to a ground station.

Satnews' February 2026 reporting on nine months of in-orbit testing frames the system as having moved from data pipes to answers over images. Europe's policy institute has now named the consequence: third-country dependence on orbital compute infrastructure. U.S. counterparts remain in prototype, with SpaceX's first orbital AI data center no earlier than 2027. Whoever reaches operational scale first owns the inference layer every other space system will rent. Shanghai's forum was institutional context. The race moved to orbit some time ago.