According to The Verge, Google has switched on the AI chips aboard its experimental satellite. The on-board Gemma 4 model then wrote a short haiku about being in space and sent it back to Earth, which was the first test of running a language model in orbit.
Key takeaways
- Google says it has powered up Tensor Processing Units on a satellite and used them to run a language model in orbit.
- The first output sent back to Earth was a haiku about being in space, produced by Google’s Gemma 4 model.
- The test belongs to Project Suncatcher, Google’s effort to find out whether AI computing could one day be done in space instead of in data centres on the ground.
- Many practical details, such as how fast the model ran, how much power it used and how long the hardware will last, have not been made public.
The satellite’s Tensor Processing Units
The main news is that the hardware works. The Verge reports that Google said in a blog post on Friday that it had booted up the Tensor Processing Units, or TPUs, on board its satellite. TPUs are chips Google designed for the maths behind machine learning. They mainly multiply large grids of numbers, which is what training and running neural networks requires. On Earth, Google uses them in its data centres to power its own AI products and to rent computing power to other customers.
Running a chip like this in orbit is much harder than running it in a server rack. The first problem is radiation. Without the protection of the atmosphere and Earth’s magnetic field, high-energy particles can flip bits in memory or damage the chip over time. The second is heat. On the ground, servers are cooled by air or liquid, but in a vacuum heat can only leave by radiating away from the spacecraft, so cooling has to be designed in from the start. The third is power. A satellite has to generate all its own electricity, usually from solar panels, and AI chips use a lot of it.
Google has not said how it dealt with each of these problems on this satellite, or how the chips are performing. Getting the processors to boot and run a model shows that the basic setup works. It does not tell us how well they would cope with months or years of constant use.
The Gemma 4 model and its haiku
The second part of the announcement is the software and what it produced. The Verge reports that the satellite ran Gemma 4, part of Google’s Gemma family of models. Google offers Gemma as a lighter, openly available alternative to its larger Gemini systems. Smaller models suit a satellite better because they need less memory, less power and less processing to give an answer.
The test was a simple prompt: write a haiku about being in space. A haiku is a short poem form from Japan, usually three lines with a set pattern of syllables. In the version The Verge quotes, the poem starts by looking down at the blue planet below, then mentions streams of data flowing in. The rest of the poem is cut off in the source material, so it is not reproduced here.
A poem is an odd way to test serious engineering, but the choice makes sense. A haiku is short, so it is cheap to transmit and easy to check. It also shows the whole chain working: a prompt sent up to the satellite, a model running on the chips there, and the answer coming back down. It makes a demonstration that is hard to picture easy to share, much as early spacecraft were known for their first photographs or first broadcasts.
The haiku says little about the model’s abilities, though. Language models write short poems easily, and one poem cannot show how the system would cope with longer or harder tasks. Google has not said how long the answer took, whether other prompts were tried, or how the result compares with the same model running on the ground.
Project Suncatcher
The address of The Verge’s article names the wider programme behind the test: Project Suncatcher. The idea is to find out whether computing for artificial intelligence could be moved off Earth, using satellites that run on sunlight and carry processors.
The reasoning starts with a problem on the ground. Large AI systems need data centres that use a great deal of electricity and water and take up a lot of land. Supporters of computing in space argue that solar panels in orbit can get steadier sunlight than panels on Earth, with no clouds and, in some orbits, little or no night. If that power could feed processors directly in space, some of the strain on electricity grids and local resources might be reduced.
The obstacles are just as large. Launching hardware is expensive. Broken components cannot easily be fixed or replaced. Moving large amounts of data between orbit and the ground needs reliable, high-capacity communication links. Doing AI work at scale would also probably mean many satellites working together, linked closely enough to share tasks. Google has not said how many satellites the project might eventually involve, what it would cost or when any working system might exist. The current satellite is best described as a test, not a service.
What the first orbital response tells us
Together, these pieces show a single experiment working from start to finish. The chips started up, the model ran and the answer came back. This is a first step, the kind of result that comes before the harder work of showing something is reliable, efficient and affordable.
It also shows where the technology industry’s ambitions are heading. Demand for computing power keeps growing, so companies are looking at unusual places to build it, including orbit. Computing in space has been discussed for years. A working test with a mainstream AI model makes that discussion more concrete, even though big questions remain unanswered.
Choosing a haiku says something about how such projects are presented to the public. A test result is hard to get excited about, but a poem from space is easy to share and remember. That makes the announcement appealing well beyond engineering circles. Science fiction has long imagined machines in orbit with something to say, and here is a small real example.
The useful information is still missing. It is not known how much computing work the satellite can actually do, how its chips will hold up against radiation over time, how much energy each task uses, or whether space-based computing could ever compete on cost with data centres on the ground. Those answers will decide whether Project Suncatcher becomes real infrastructure or stays a notable experiment. For now, the evidence is limited to a short poem and a confirmation that the hardware works.
Sources and further reading
- The Verge: technology coverage reporting Google’s announcement of the satellite’s first AI-generated response.
- Google: the company’s blog post describing the TPU boot-up and the Gemma 4 haiku test, as cited by The Verge.
- Google’s public material on Gemma: background on the company’s family of openly available language models.
- General space engineering references: background on radiation, cooling in a vacuum and solar power for spacecraft electronics.
Surfaced from the rss:verge signal “AI model runs in orbit”. AI-assisted draft, editorially reviewed.
