Google’s TPU in orbit answers a prompt with a haiku about space

Google’s Project Suncatcher prototype satellite has returned its first AI response from orbit. Sundar Pichai, Google’s CEO, called it a “hello world” milestone: the team sent a command up, booted the TPU, ran a prompt and received the answer.

Google Research gave the details in a blog post by Travis Beals, a senior director at the company. The prompt asked the open Gemma 4 model running on the satellite for a haiku about being in space. The poem came down as the satellite passed over Svalbard, Norway.

A haiku from a TPU circling Earth

According to the post, the satellite launched on October 1 and the team then worked through a commissioning checklist over several weeks. That included powering up the host and the TPU trays and running early machine learning tests. Google describes the haiku as the first AI response from a data center class TPU in orbit.

The poem read: “Blue marble below / Data streams in silent grace / Starlight feeds my core.” Follow-up tests used space-themed limericks and telemetry prompts in which the model reported the status of the payload systems. In the post, Google says its chips “have officially earned their space wings.”

The post does not say how long the response took, how many tokens were generated, or how much power the TPUs drew. It gives no orbit details either, beyond the Svalbard ground-station pass.

What the launch and hardware look like

Public reports at the time of launch fill in some of the gap. Several NPR member stations, including Georgia Public Broadcasting, reported that the prototype flew on SpaceX’s Transporter-18 rideshare mission from Vandenberg Space Force Base in California. They said it carries four TPUs and was built with Planet. The same coverage noted that the prototype sits in a standard orbit and runs on batteries, and that the laser links planned for later satellites are not part of this one. Google’s new post does not repeat these points, so they rest on that earlier coverage.

Limits and open questions after one poem

A haiku proves that the chip booted and ran an inference job in orbit. It does not show that the hardware holds up in the long run. Google says the next tests will look at how the TPUs cope with thermal stress and radiation exposure. Its original Suncatcher post said initial radiation testing of the chips had begun, but it did not publish results in the text we read.

The timeline also shifted. Reporting on the launch said Google’s November 2025 roadmap had put the first prototypes in the early months of 2027, so an October 2026 launch is earlier than planned. Google’s stated next milestone is a pair of interconnected satellites in 2027, and the earlier post said that learning mission is being run with Planet. That step matters more than the haiku, because the concept depends on many satellites exchanging data with each other, not on a single chip answering a prompt.

The economics are untested too. A prototype that runs on batteries says little about the solar-powered design Google describes for the full system, and the post offers no figures on cost, heat rejection or power per satellite. Google presents the project as a long-term research moonshot, not a product, and nothing in the announcement suggests customers will be able to rent compute in orbit soon. For readers following the AI infrastructure race, the signal is that Google is willing to test whether space can supplement ground data centers, which are increasingly constrained by power. The evidence so far is one successful prompt.