Google Sent a TPU to Space, and the First Task It Ran Had AI Write a Haiku

Image credit: Google

Google's AI satellite has sent back the first model response ever generated directly in outer space. Google announced in a blog post that it successfully activated the Tensor Processing Unit (TPU) on the satellite, and ran a prompt on the in-orbit Gemma 4 model to write a haiku about being in space:

Blue marble below
Data streams in silent grace
Starlight feeds my core. ✨

Google also tested other space-themed limerick prompts and telemetry prompts for the hardware. The test falls under Project Suncatcher, an experimental program with a long-term vision of deploying full AI data centers in space. Google launched the satellite into orbit earlier this year.

This haiku does not prove that space-based data centers are already practical, of course. What it confirms is a much narrower milestone: the AI chip powered up successfully in orbit, the model completed full inference directly on the satellite, and the results were transmitted back to the ground. For Project Suncatcher, this is a necessary step, and also the most publicly verifiable milestone achieved to date.

Supporters of space-based data centers highlight the advantages of abundant solar energy and superior heat dissipation conditions in orbit. Critics first raise concerns around maintenance, radiation interference, launch costs and communication bandwidth. Google did not address any of these large core questions with this test, it only demonstrated that "a TPU can boot up and run AI models in space". The first output being a haiku is very on-brand for Google, and looks exactly like a carefully controlled demonstration.

Project Suncatcher is still in its early phase. Two points are worth keeping in mind when reading this news: Gemma 4 did generate the text truly in orbit, and the TPU was successfully activated; there is still a huge engineering gap between this single satellite and a full "space-based AI data center".

Source: Google Official Blog、The Verge