Hyderabad: Google has achieved a landmark moment in its ambition to run Artificial Intelligence (AI) from space. The Project Suncatcher prototype satellite, launched on October 1st 2026 aboard a SpaceX Falcon 9 rocket alongside more than 100 other spacecraft, has successfully generated an AI response from orbit and beamed it back to Earth. It is the first time a data centre-class AI chip has operated in space. 

What is Project Suncatcher?

  Project Suncatcher is Google's long-term research project exploring whether AI computing can be run from satellites in orbit rather than from energy-intensive data centres on the ground. If it works at scale, it could reshape how global AI infrastructure is built and powered.

The prototype satellite carries four of Google's own Tensor Processing Units (TPUs), the specialist chips the company uses to run AI models. However, the power demands of a TPU are significant. The satellite generates up to one kilowatt of electricity, which means it can only run one TPU at a time, for around 20 minutes per pass. This places real constraints on how much work can be done during each contact with a ground station. 

Three weeks of testing before the main event

After launch, Google spent approximately three weeks completing a commissioning process, powering up onboard systems, testing subsystems, and running initial machine learning tasks. Once satisfied that the satellite was functioning correctly, the tech giant was ready to send its first real prompt.

The Hello World moment

During a ground-station pass over Svalbard in Norway, engineers sent the first-ever prompt to an AI model running onboard a satellite in orbit. The model was Gemma 4, an open AI model from Google DeepMind. The team asked it to write a haiku — an unrhymed short-form poem — about being in space.

According to TechCrunch, a spokesperson from the company confirmed that the satellite was operating normally. 

The radiation problem

One of the biggest unknowns for AI chips in space is radiation. In orbit, cosmic radiation can cause bitflips, which are tiny errors in computer memory that can corrupt data without triggering standard monitoring tools. Google has tested its TPUs in particle accelerators on Earth to simulate this effect, but the real-world conditions of space are different, and Google now needs to assess how radiation actually affects performance in orbit.

Early ground testing suggested radiation would have a minimal impact on inference workloads — running an existing AI model — but would be more of a problem for training a model from scratch. The ongoing test campaign will now put those findings to the test in reality.