Google's test satellite answers an AI prompt from orbit, sending back a haiku

A haiku from orbit

Google has run an AI model on a satellite circling the Earth and received a full answer back on the ground. As The Economic Times reported, the test is part of Project Suncatcher, a long-term research effort asking whether space could host the computers that power machine learning.

Google CEO Sundar Pichai announced the result on X and called it a "hello world" milestone. That phrase is programmers' shorthand for the first, simplest program that shows a new system works at all.

The team sent a command to the satellite, switched on its Tensor Processing Unit (TPU), ran a prompt and got the output back. A TPU is a chip Google makes specifically for AI workloads. The prompt asked for a haiku about space, and the answer was Pichai's headline detail: the first complete response, a haiku.

How the test unfolded

According to Google's blog post, the prototype satellite launched on October 1. In the weeks after, the team worked through commissioning checks, the routine step where operators confirm a new spacecraft's systems behave as expected. That included powering up the satellite's host computer and its TPU trays, and running early machine learning tests.

Then came the prompt. During a pass over a ground station, the team sent the request to Gemma 4, an open model running on board. An open model is one whose underlying files are shared publicly, so others can run it. The reply came down as the satellite passed over Svalbard, Norway.

Google says it kept going afterward. It tried space-themed limericks, and it asked the model to report on the technical status of the satellite's own systems.

Why put AI computing in space at all?

The pitch rests on sunlight. Google has said that satellites in low Earth orbit, meaning a path relatively close to the planet, can get near-constant sun. In its telling, they could produce up to eight times more solar power than panels on the ground.

That matters because running large AI systems takes a lot of electricity. If power could be gathered more steadily in orbit, the thinking goes, some AI work might eventually be done there.

Suncatcher also looks beyond a single satellite. The project is exploring whether satellites could be linked into constellations, groups that work together, to handle bigger AI jobs. That is a research question, not a plan with a price tag or a launch schedule for customers.

The harder tests are still ahead

Answering a haiku prompt shows the chip can start up and run a model in orbit. It does not show the chip can keep doing so for long stretches. Google says the next phase will test how the TPUs cope with thermal stress, meaning big swings in heat, and with radiation exposure in orbit.

Those tests are meant to settle whether the chips can reliably handle computing work in the space environment. Google says that if they succeed, the next milestone is two interconnected satellites in 2027.

• The first step, running a model and getting an answer back, is done.
• The second step, showing the hardware holds up over time, is under way.
• The third step, linking satellites together, is planned for 2027.

How much to read into it

It is easy to oversell a poem from orbit. The coverage came with few details, and the source says nothing about costs, how much computing the satellite can do, or when anything like this could be useful at scale. This is one prototype passing its first exam.

Still, the result is real and specific. A company that builds its own AI chips has shown one working in orbit, answering a request sent from the ground. Whether the idea becomes a practical way to power AI will depend on the heat and radiation results and the 2027 two-satellite test.