Google’s first data center satellite is online and writing poetry

Google's Suncatcher project says its first data center satellite is online and communicating with the ground, after launching earlier this month on a SpaceX Falcon 9 alongside more than 100 other spacecraft. The prototype carries a single Google Tensor Processing Unit, and its debut task was a prompt from engineers asking the on-board model to compose a haiku about being in space.

The spacecraft runs Google DeepMind's Gemma 4, pre-loaded before launch. Google also had the model produce space-themed limericks and, in a more practical exercise, asked it to report on the orbital conditions around the satellite. Power is the binding constraint: the TPU's electricity budget means it can only be switched on for roughly twenty minutes at a stretch, which complicates the already tricky business of space-to-ground communications.

A commissioning process tested the satellite's subsystems before the first prompt went up. The ongoing test campaign is meant to show how the chips handle thermal stress and the space radiation environment. Google has put TPUs through particle accelerators on Earth, but radiation in orbit can cause bit flips in stored memory, corrupting data without triggering typical monitoring tools. Ground testing suggested radiation would have minimal impact on inference and be more of a problem for training — the team can now assess those dynamics in reality rather than simulation.

If the orbiting TPU proves capable, Google's next step is launching two more satellites equipped with advanced chips in 2027, carrying higher power generation and demonstrating inter-satellite optical links. Suncatcher is largely alone among space-based data center visions in planning high-speed links between close-flying satellites to support coherent distributed workloads. SpaceX's data centers are also expected to use laser links, but their satellites will orbit too far apart to share processing effectively, while companies like Starcloud and Cowboy Space expect to run inference loads on single spacecraft.

Both approaches lean on Starship, the still-in-development heavy-lift rocket, becoming fully and rapidly reusable; a peer-reviewed analysis from Suncatcher researchers estimated SpaceX would need to fly Starship 1,800 times to achieve the learning-driven cost reductions it saw with earlier vehicles. Reaching that point may depend on nearer-term economics: Alphabet's stock plunged this week after the company reported that inserting AI into search is hurting ad revenue growth. Open questions remain whether the TPU survives real radiation across repeated power cycles, whether twenty-minute operating windows can support meaningful inference workloads, and how long a 2030s-oriented moonshot stays funded while core search revenue is under pressure.