Google's AI Chip in Orbit Has Written Its First Haiku
Google says a satellite carrying its AI chips has sent back its first response from space. The AI model on board, called Gemma 4, wrote a haiku about being in space after receiving a prompt. The Verge reported the transmission on October 30, 2026.
The haiku reads: "Blue marble below / Data streams in silent grace / Starlight feeds my core."
The poem is a small test, but it covers a complete cycle. A question was sent up to the satellite, the AI model ran on hardware in orbit, and the answer came back down to Earth.
The chips are Google's Tensor Processing Units, or TPUs, which are built to run AI software. Google says it successfully switched on the satellite's TPUs in orbit. The team also tested other space-themed limericks and prompts that check the satellite's own status readings. The company's blog post on the satellite response describes the same sequence of tests.
The work falls under Project Suncatcher. Google calls it an experiment with a long-term goal of putting AI data centers in space. Google Research first described the idea in a November 2025 post, "Exploring a space-based, scalable AI infrastructure system design". That post says the plan is to equip solar-powered groups of satellites with TPUs. The first test satellite is a joint project between Google and Planet, according to Google's September 2026 AI updates, published October 2, 2026.
Google plans more tests to see how its TPUs handle heat and radiation in orbit. These two factors are the biggest differences between a chip on the ground and one in space. On Earth, air and liquid carry heat away from chips, and radiation rarely troubles ordinary computer parts. In low Earth orbit, there is no air. Heat has to leave as infrared radiation, much like a warm stone giving off heat into a cold room. High-energy particles can also flip stored bits of data, so engineers have to measure that effect instead of assuming it is small.
The next goal, once TPUs have proven reliable in space, is a pair of satellites in 2027 that communicate with each other. Two linked satellites would be a first step toward spreading computing across several machines instead of one. Google has not yet said how the two would connect, how much data they could move, or which orbits they would use.
In my view, the haiku is the detail most people will remember, and it is worth remembering. But it is also the simplest part of the test. One short text reply says little about sustained performance, how memory holds up under radiation, or whether temperatures stay steady across a full orbit. Those measurements will decide whether Suncatcher becomes a real working system.
The broader context here is that AI computing on Earth is increasingly limited by power and cooling. The Suncatcher idea is that solar power in orbit could ease part of that limit. Whether the idea holds depends on the data the coming tests are designed to collect. For now, the confirmed result is narrow: a model running on TPUs in orbit sent a reply that reached the ground. The next results will come from the heat and radiation tests, and from the 2027 two-satellite milestone.
