Google’s experimental Suncatcher satellite has executed an AI model in orbit and returned a complete response to Earth, a milestone the company is framing as the first AI output generated by a data-centre-class chip in space. The prototype, launched on 1 October, ran Google’s open Gemma 4 model on its onboard Tensor Processing Unit and downlinked a three-line haiku as it passed over Svalbard, Norway. Travis Beals, a senior director at Google, announced the result in a blog post, calling it a symbolic ‘hello world’ moment for orbital computing.
- First AI response from orbit: A ground station uplinked a prompt to Gemma 4, and the satellite returned a haiku about being in space, confirming the full command-process-downlink loop.
- Hardware constraints: The satellite carries four TPUs but generates only up to 1 kilowatt of power, limiting it to running a single chip for roughly 20 minutes at a time.
- Next phase: Google will test how the TPUs withstand thermal stress and radiation in orbit, with plans to launch two linked satellites in 2027 if the chips hold up.
What the satellite actually did
During a pass over a ground station, engineers sent a prompt to the onboard Gemma 4 model asking it to write a haiku about being in space. The response came back as the satellite flew over Svalbard: ‘Blue marble below / Data streams in silent grace / Starlight feeds my core.’ The team then tried space-themed limericks and prompts asking the payload to report on its own technical status. A Google spokesperson told TechCrunch the satellite is operating nominally after roughly three weeks of commissioning.
The claim is carefully worded: the chip is the same class of hardware that powers Google’s terrestrial AI systems, not a smaller, space-hardened processor of the kind already used on many spacecraft. Omar Sanseviero, who works on Gemma at Google DeepMind, noted on X that the model ran on stored solar energy. The satellite was built in partnership with Planet and launched aboard a SpaceX rideshare flight.
Sundar Pichai described the result as a ‘hello world’ milestone for Project Suncatcher, writing that the team had ‘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!’
Why Google wants AI in space
Project Suncatcher, first outlined by Google in November 2025, asks whether constellations of satellites flying in close formation in dawn-dusk sun-synchronous orbits could offer near-constant sunlight for solar power and free-space optical links fast enough to move AI workloads between spacecraft. In Google’s published design, clusters would be connected by laser links running at tens of terabits per second, allowing work to be shared across a formation rather than done on a single spacecraft.
The appeal is energy. A satellite in the right orbit can see the sun almost continuously, avoiding the night-time gaps that limit ground-based solar power, while space itself provides natural cooling. The obstacles are equally clear: radiation, heat, launch cost and the difficulty of repairing hardware that cannot be reached. Google’s ground tests, published in the journal Joule, described proton-beam testing of the Trillium TPU in which high-bandwidth memory showed irregularities only after a cumulative radiation dose of 2 kilorads, a level the team presents as comfortably above its minimum requirement.
The orbital mission is the next step: finding out whether those laboratory results hold up in the real radiation and thermal environment of space. Radiation can flip bits in a chip’s memory and corrupt data without warning, though Google’s ground tests suggested this matters less for running models than for training them.
The hard part is still ahead
It is important not to over-read a haiku. The test shows that a TPU survived launch, powered up in orbit, ran a model loaded before launch and returned a complete answer. It does not show speed, cost, reliability over months or any advantage over running the same model on Earth. Google has published no benchmark figures from the orbital run, and the company is explicit that Suncatcher remains a research experiment, not a product.
The 20-minute operating window is the detail to watch. If a chip can only run in short bursts, any future space-based system would need to schedule work carefully around power, heat and ground-station contact, or generate far more power per satellite. Google’s next planned step addresses that: two linked satellites in 2027 that would make more power and test laser links between them. The plan also depends on SpaceX making its Starship rocket fully reusable, because without cheap launches the economics do not close. SpaceX has told investors its own orbital data centres may never work.
Google is treating space as the next frontier of AI infrastructure, not for spectacle but because terrestrial data centres are running into the limits of power grids, water and land. The experiment has succeeded at its easy part: booting a chip and receiving a reply. The difficult part — radiation, heat, the cost of launching tonnes of hardware and keeping it running — remains unresolved. For now, Suncatcher is an expensive proof of concept that has written a rather nice haiku.