Google says its Project Suncatcher satellite has returned the first AI response from a TPU in orbit. The company reported the milestone on October 30 S¹. The company uplinked a prompt to a Gemma 4 open model running on the TPU aboard the prototype launched October 1 S¹. The TPU is Google's custom AI chip. Google labels the milestone as the first time a data center class TPU has produced an AI response from orbit. It is a self-reported claim no one outside the company has verified S¹.
My read: The haiku is a charming demo, not a workload. What matters is whether the TPU survives thermal stress and radiation over weeks and months, and Google has released zero performance data so far. The 2027 twin-satellite milestone is a stated plan, not a funded commitment. I would watch for whether the next round of tests produces actual telemetry on chip degradation, not more poetry.
Google launched the prototype on October 1, built in partnership with Planet aboard SpaceX's Transporter-18 rideshare mission P². Travis Beals, Senior Director of Paradigms of Intelligence at Google, confirmed contact with the satellite shortly after launch P². Over the following weeks, the team completed a commissioning checklist that included powering up the host and TPU trays and running early machine learning tests S¹.
How the first response worked
During a ground-station pass, the satellite was within radio range of a ground antenna. Google sent a prompt up to the Gemma 4 open model on board, requesting a haiku about life in space S¹. When the satellite passed over Svalbard, Norway, Google received the poem back, confirming the round trip worked S¹.
The haiku read: "Blue marble below / Data streams in silent grace / Starlight feeds my core." S¹
Google then tried space-themed limericks and telemetry prompts that checked the status of the payload systems S¹. These are lightweight natural language tasks, not the kind of compute that would stress a data center class chip, a term Google defines itself without external technical specification S¹. Google has not published any performance metrics, power output data, or error rates from the TPU in orbit.
The tests that actually matter
The next phase will test how the chips hold up under orbital thermal stress and radiation S¹. Those results, when and if they come, will tell whether orbital AI compute is viable or just a headline.
The company's next stated milestone is twin interconnected satellites in 2027 S¹, part of a research moonshot first announced in November 2025 to explore scaling machine learning compute in space P³. For satellite operators and edge-compute engineers, the signal to watch is whether Google publishes raw telemetry from those thermal and radiation tests. A haiku proves the communication link works. It says nothing about whether a TPU can run a real model workload for months without degrading.
Google's chips, in the company's own words, have "earned their space wings" S¹. Whether they keep flying is a question for the next test cycle.
Sources: S1 — We’ve received our first response from the Project Suncatcher prototyp · P2 — Google's Project Suncatcher prototype satellite is in orbit · P3 — Project Suncatcher explores powering AI in space · P4 — google (Google) · P5 — leachim6/hello-world
Written from 5 sourced items, 5 of them primary.
