We are used to it Artificial intelligence For everyday tasks like solving a question or translating a text. We do this from our mobile phones or computers, without relying too much on the massive computing resources that make each response possible. However, behind that apparent simplicity lies a physical infrastructure that requires power, cooling and communications to function. And as we continue to expand its capabilities here on Earth, a possibility that until recently seemed far-fetched is also being explored: taking some of that computing into space.
Google has already taken a first step to explore that possibility. On October 30, the company announced that it had received a response from orbit generated by its Gemma 4 open model using a Tensor Processing Unit (TPU), a specialized data center-grade chip. The satellite, built in collaboration with Planet and launched Oct. 1 on a SpaceX mission, is part of Project Suncatcher, an initiative that investigates space computing. For now, the result is not equivalent to getting a data center up and running. The test consisted of something very simple: asking the model to write a haiku about the place.
To check if everything was working, the researchers took advantage of a communication window with a ground station and sent a command to the satellite: Write a haiku about being in space. It is a very short poem of Japanese origin, usually consisting of three stanzas. The answer came as the instrument approached Svalbard, Norway, and translated from English: “There’s the blue sphere,” “Data flows with quiet elegance” and “Starlight feeds my core.” Beyond poetry, the important thing was to verify that the system could process the request in orbit and return the result to Earth.
From haiku to the challenge of computing in space
The test also highlights Google’s longstanding technological commitment. In 2016, the company unveiled its TPUs, processors designed to accelerate artificial intelligence calculations that it had already used internally since 2015. With them it has an alternative to NVIDIA GPUs, although this does not mean complete supply chain independence: Broadcom participates in its development and participates in its management. The prototype includes four TPUs, but can only use one at a time for a short period of around 20 minutes, according to TechCrunch. This is a testbed, not a full data center.
There is an important nuance up front: this is not the first time that artificial intelligence has been implemented in orbit. In 2020, the European Space Agency’s Φsat-1 mission used an AI system to filter cloud-covered images of Earth before transmitting them. In November 2025, Starcloud launched a satellite that was a NVIDIA H100 GPU And the following month announced that it had launched a Google model called Gemma on board. The innovation the company is now claiming is more specific: getting a response through a data center category TPU in space.
Project Suncatcher’s ambitions, however, extend far beyond this first experiment. Google plans to build a distributed infrastructure to run artificial intelligence tasks, made up of satellites equipped with processors and connected to each other through optical links. The big advantage it seeks is power: in certain orbits, near-permanent exposure to the sun would allow it to generate up to eight times the electricity that solar panels on Earth would generate, he estimates. If achieved, it would have a new way to expand computing power, although it is still far from proving that it is viable on a large scale.
The problem is that space conditions also pose very different difficulties than those faced by a terrestrial data center. However it may seem Cold cold is easierIn a vacuum there is no air to transport the heat generated by the processor. This is why Google is investigating systems with thermal ducts and radiators, which it has already tested in a vacuum chamber. Added to this is space radiation, which is capable of altering the operation of circuits and against which it has also conducted ground tests. You still need to test how your hardware performs in the long run and whether it is cost effective to deploy.
After the haiku, the researchers tried other space-themed poems and requests aimed at learning the technical status of the satellite. They are now preparing tests to observe how the TPU withstands thermal stress and radiation while operating in orbit. The next announced goal is to test two interconnected satellites in 2027, a step to study how these systems can work together. That first answer therefore does not confirm that a spatial data center is viable. This confirms something more limited, but significant: that a fundamental part of the concept is already able to respond from space.
Picture | Google
In Xataka | Sam Altman, CEO of ChatGPT: “It’s scary that people’s PCs only have RAM 5% of the time being used”
