Google puts a TPU in orbit and it wrote a haiku

While tech giants on Earth are scrambling to lock down nuclear plants and power grids to feed their hungry data centers, Google is exploring the ultimate off-grid alternative: outer space.

Google announced that its Project Suncatcher prototype satellite successfully executed its first AI inference in orbit. The milestone marks the first time an AI response has ever been generated in space using a data center-grade TPU.

A haiku beamed down from Norway

The prototype satellite actually reached orbit earlier this month on October 1. Over the past few weeks, Google engineers have been working through a commissioning checklist, powering up the onboard host computers, initializing the TPU trays, and running baseline machine learning workloads.

This week, the orbital rig hit its "Hello World" moment. As the spacecraft passed over a ground station near Svalbard, Norway, the team uplinked a prompt asking the onboard Gemma 4 open model to compose a haiku about being in space.

The satellite downlinked three poetic lines to confirm success: "Blue marble below / Data streams in silent grace / Starlight feeds my core."

Beyond poetry, Google followed up with space-themed limericks and practical telemetry prompts designed to report the technical health of the payload systems. The chip officially earned its space wings, as Google put it.

The hard physics still lie ahead

Getting a server chip to boot up and spit out a poem in orbit is neat, but the brutal realities of space flight are just beginning to kick in.

In a terrestrial data center, massive cooling loops and fans keep high-wattage silicon happy. In the vacuum of low Earth orbit, there is no air for convection. Chips face violent thermal swings between sunlight and shadow, alongside constant exposure to ionizing cosmic radiation that can flip bits or fry circuits.

Google says its immediate next step is running the TPUs through extensive thermal stress and radiation testing. If the silicon proves it can reliably chew through workloads under these hostile conditions, Google plans to launch twin interconnected satellites in 2027 to test orbital clustering.

Why it matters

For everyday developers, orbital TPUs won't change your daily workflow anytime soon, and beaming down poetry is undeniably a bit of a PR spectacle. But for aerospace and remote sensing, onboard frontier models could be massive. Right now, Earth observation satellites spend precious downlink bandwidth beaming raw pixels back home. Running models like Gemma 4 in orbit means satellites can analyze, filter, and summarize sensor data directly in the sky.

For the broader AI industry, this opens a fascinating new flank in the infrastructure race. As land-based power constraints tighten, Microsoft and Amazon are betting heavily on nuclear and geothermal power, while Google is prototyping space-based solar computing. If orbital data centers ever become viable, compute competition will literally transcend Earth's atmosphere.

We think Google's approach here is remarkably smart. Instead of building exotic, unproven space-only chips, they strapped production-grade TPU hardware and an open-weights model onto a satellite to see how commercial hardware behaves. Radiative cooling in a vacuum remains an enormous engineering hurdle, and the 2027 twin satellite mission is years away, but proving that standard enterprise AI silicon can survive launch and run inference in orbit is a genuinely impressive first step.