Google's in-orbit TPU returns first AI response as Gemma 4 writes haiku in space

Google launched a Project Suncatcher prototype satellite on October 1, 2026, local time. Over the following weeks, the team worked through an on-orbit test and commissioning checklist that included powering up the host and TPU trays and completing early machine-learning tests. On October 30, local time, Google announced that it had reached a “Hello World” milestone: a datacenter-class TPU running in orbit produced its first AI response. That week, engineers on the ground sent the first prompt and asked Gemma 4, a Google DeepMind model preinstalled on the satellite, to write a haiku about being in space. The model replied: “Blue marble below, Data streams in silent grace, Starlight feeds my core.” IT之家 provided a Chinese translation: 俯瞰蔚蓝丸，数据无声流淌去，星光赋机芯. Beyond the haiku, engineers also tested five-line limerick generation and telemetry tasks such as returning orbital environment data.

The satellite carries four Google TPUs and can generate up to 1 kW of power. Google said thermal limits and energy load constrain the TPUs so that they can be switched on for only about 20 minutes at a time. The prototype is described as laying groundwork for a future tightly-formed satellite network intended to run large-scale inference workloads in orbit.

Reports on October 31 said Google’s first datacenter satellite was now online, with the first TPU in space running Gemma 4; that report did not disclose the scale of the space compute or later deployment plans. On the same day, a Reddit post citing a Google blog stated that Google confirmed the Project Suncatcher satellite had successfully run Gemma 4 on an onboard TPU and transmitted the first generated response back to Earth, which was a poem. As of October 31, public information remained at Google’s milestone announcement, the haiku and other telemetry tests, and the satellite’s limited four-TPU power and roughly 20-minute operating window per activation.