A Gemma 4 model running on a Google TPU in low Earth orbit completed its first inference task and beamed a poem back to Earth
Google’s Project Suncatcher satellite ran a prompt through its onboard AI model and sent the result home as a haiku.
The event marks the first time a data-center-class Tensor Processing Unit has run successfully in space after surviving launch.
From launch pad to poetry
The prototype lifted off on October 1, 2026. It flew aboard SpaceX’s Falcon 9 Transporter-18, a rideshare mission, under a partnership with Planet Labs.
The satellite then went through several weeks of commissioning.
Its first AI inference task came on or around October 30, 2026, during the week of October 27 to 30. Inference is the stage where a trained model actually answers a prompt, as opposed to the earlier, heavier work of learning from data.
The answer took the form of a haiku that reflects the satellite’s space surroundings. The model behind it was Gemma 4, an open model from Google DeepMind.
The hardware doing the work is a Google TPU. TPUs are Google’s custom chips built for machine learning, the same family of processors that runs AI workloads in the company’s data centers.
The constraints are very real
The setup is modest by design. The satellite runs on a single Google TPU powered by approximately 1 kW of solar energy.
Operating sessions are capped at 15 to 20 minutes because of thermal constraints.
Why Google wants data centers in orbit
Google announced Project Suncatcher in late 2025. The stated aim is to harness solar power in low Earth orbit to build machine learning infrastructure that can scale.
The next step is already sketched out. Plans for 2027 include expanding the mission to two interconnected satellites.
Those satellites are intended to handle larger workloads and to test laser communication links between spacecraft.
What this means
For Google, the milestone is primarily a hardware validation. The company now has evidence that its in-house AI chips can endure launch and run a modern open model in orbit, which is the foundational question for the entire Suncatcher concept.
The partnership structure is also worth noting. Google supplied the TPU and the model, SpaceX supplied the ride, and Planet Labs served as the mission partner.
The use of Gemma 4 carries its own signal. Running an open model rather than a closed flagship shows Google is willing to test its publicly available tooling in extreme conditions, and it gives developers a concrete example of where that model family has been deployed.
The open questions are practical ones. Short operating windows of 15 to 20 minutes, limited power of approximately 1 kW, and heat management all stand between a demonstration and anything resembling a production workload.
