Google has run an AI model on a prototype data centre satellite in orbit and received its answer on Earth. The onboard Gemma 4 model wrote a haiku about being in space, senior director Travis Beals wrote in a blog post on Friday. Google described it as the first AI response from a data centre class TPU, its own AI chip, in orbit. The test matters because it moves orbital AI computing from concept to a working inference run.
First inference run on the Suncatcher satellite
The satellite launched on 1 October for Project Suncatcher, Google's research effort to run AI computing in space. Google sent the prompt during a pass over a ground station, and the poem came back as the satellite flew over Svalbard, Norway. The published haiku reads: "Blue marble below / Data streams in silent grace / Starlight feeds my core." The team then tried space-themed limericks and prompts asking the satellite to report on its own systems.
The prompt asked for a haiku written as a TPU orbiting Earth, as reported by Thomas Claburn for The Register. The satellite holds four TPUs and generates up to 1 kilowatt of power, Tim Fernholz reported for TechCrunch. Because of that power limit, it can run only one TPU at a time, for about 20 minutes. Testing its systems took about three weeks, and a Google spokesperson told TechCrunch the satellite was working normally.
Omar Sanseviero, who works on Gemma at Google DeepMind, wrote on X that the model ran on stored solar energy. Google will now test how the TPUs cope with heat and radiation in orbit. Radiation can flip bits in a chip's memory and corrupt data without warning. Google's tests on the ground suggested this matters less for running models than for training them, according to TechCrunch.
What orbital inference means for AI operations
If the chips hold up, Google plans to launch two linked satellites in 2027. They will make more power and test laser links between satellites, TechCrunch reported. For companies using AI, that points to a possible future source of inference capacity located off Earth. A small firm would feel it only as another cloud region or vendor option, while a large operator could compare orbital capacity with ground data centres on latency, availability, and energy terms.
The limits are clear from the current prototype: one TPU at a time and around 20 minutes of operation within a 1 kilowatt envelope. Heat removal and radiation-induced memory errors remain open checks, with the next phase focused on measurement rather than commercial service. This news alone does not show cost, reliability, or scale for business workloads. Vendor questions to keep include power budget per chip, error rates during inference, downlink constraints, and how ground stations and laser links affect response times.
The marker to watch is the pair of linked satellites planned for 2027 and their laser-link tests. Successful operation with higher power would show that orbital inference can move beyond single short runs. Failure or delay would keep space data centres as research hardware. Either outcome will shape how seriously firms treat orbit as part of AI infrastructure planning.
