AI datacenters in space are suddenly not looking so far fetched any more.
Google has hit a milestone in its space-based AI computing effort, Project Suncatcher, successfully sending a command to its prototype satellite in orbit, booting up the onboard TPU, running a prompt, and receiving a complete AI-generated response back on Earth.
Announcing the update, Google CEO Sundar Pichai called it a “‘hello world’ milestone for Project Suncatcher,” noting that after launching the prototype satellite a few weeks ago, the team “successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth – a haiku!”
On being asked to write a Haiku from its perspective, the TPU sent back:
Blue marble below,
Data streams in silent grace,
Starlight feeds my core.
The milestone was reached during a ground-station pass, when the team uplinked a prompt to the onboard Gemma model, asking it to generate a haiku about being in space. As the satellite passed over Svalbard, Norway, the downlinked response confirmed the test had worked: a short poem describing the view of Earth from orbit and starlight powering its systems. Following the haiku, the team also tested space-themed limericks and prompts asking the satellite to report telemetry on its payload systems.
Project Suncatcher’s prototype satellite, built in partnership with Planet, launched on October 1 aboard SpaceX’s Transporter-18 rideshare mission, carrying four Trillium-generation TPU chips into low-Earth orbit. In the weeks since, the team has worked through commissioning checks, powered up the host and TPU trays, and run initial machine learning tests before reaching this “hello world” point. The next phase of testing will look at how the TPUs hold up under thermal stress and radiation exposure over longer stretches in orbit, ahead of a planned 2027 launch of two satellites designed to test high-bandwidth laser links between spacecraft.
Project Suncatcher is Google’s long-term bet on moving AI compute off the planet entirely. The idea is to build compact constellations of solar-powered satellites, each carrying TPUs and linked by free-space optical communication, so that AI workloads can eventually run on a data center suspended in orbit rather than on the ground. The rationale, as Pichai has explained previously, comes down to energy: the sun puts out roughly 100 trillion times more power than humanity’s total electricity production, and satellites in sun-synchronous orbit can draw on that supply nearly around the clock, something no ground-based data center can match.
Google isn’t the only one racing to get AI chips working in orbit. Nvidia-backed startup Starcloud beat Google to a similar milestone last year, flying an Nvidia H100 GPU aboard its Starcloud-1 satellite in November 2025 and using it to train an AI model in space for the first time, running Andrej Karpathy’s nanoGPT on the complete works of Shakespeare as well as inference on a preloaded Gemma model. Starcloud has since raised hundreds of millions of dollars to scale up its orbital data center ambitions, with Nvidia itself joining as an investor, and is working toward satellites carrying multiple H100s and, eventually, Nvidia’s newer Blackwell-based hardware.
For now, Google’s own in-space compute remains modest by design, four TPUs running in short bursts to validate that the hardware survives launch, radiation, and the thermal swings of orbit. But the company has been explicit that it sees this as an early step toward something much larger: a network of TPU-equipped satellites that could one day supplement, or even replace, the energy-hungry data centers AI companies currently have to build on Earth.