# Google Suncatcher Runs Gemma 4 in Space: What the First AI Response Proves

Google has received an AI-generated response from orbit, but sustained computing and satellite-to-satellite links are the harder tests ahead.

By Paras Madan | 2026-10-31

Google said on October 30 that Project Suncatcher ran Gemma 4 in orbit and returned a response to Earth. The test advances its space-computing experiment, without establishing that orbital AI infrastructure is ready for commercial workloads.

The interesting part is the complete engineering chain, rather than the literary quality of a model response.

## What Google tested in orbit

According to [Google's announcement](https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-satellite-response/), the team powered the host and TPU trays, then tested machine learning. A prompt sent to the onboard Gemma 4 model produced a poem, received over Svalbard, Norway.

Google calls this the first AI response from a data-center-class TPU in orbit, a hardware-specific claim rather than a claim about all satellite AI.

## What changed since the October launch

The prototype [launched on October 1](https://blog.google/innovation-and-ai/models-and-research/google-research/project-suncatcher-prototype/) aboard SpaceX's Transporter-18 mission, in partnership with Planet. At that point, Google had confirmed contact and expected operations. The new update adds completed model execution and a returned result.

These are different engineering checkpoints. A satellite can communicate successfully while an experimental payload still needs testing. Running a model closes another part of the chain, but repeated operation matters more than one successful exchange when the eventual ambition is infrastructure.

## Why sunlight does not solve the cooling problem

Google's [technical overview](https://blog.google/innovation-and-ai/models-and-research/google-research/google-project-suncatcher-facts/) explains the appeal of near-constant sunlight and the difficulty of removing heat in a vacuum. There is no airflow to carry heat away. The team is testing cooling approaches involving heat pipes and radiators.

That makes the phrase “AI in space” easy to misunderstand. Moving a processor away from Earth's electricity grid does not remove the rest of its operating requirements. Power collection, heat rejection and reliable communication still have to work together, under conditions very different from a server building.

For AI operators, the meaningful question is sustained useful compute. A machine that can complete a short request has demonstrated a capability. A machine that can repeatedly complete valuable workloads within a predictable operating envelope begins to look like capacity someone could plan around.

## The next test is reliability, then connected satellites

Google says it will test TPU performance under thermal stress and radiation exposure, with twin interconnected satellites planned for 2027. Its earlier overview describes laser communication between future satellites, using high bandwidth over short distances to support larger workloads.

A connected system introduces another dependency: compute is useful only when information reaches the next processor when required. For a distributed workload, adding chips is not enough if communication becomes the bottleneck. The 2027 milestone therefore tests a different part of the architecture from the single-satellite result.

## What this means for AI infrastructure

The announcement does not provide customer access or pricing. A commercial comparison would need much more: the useful work delivered over the hardware's lifetime, the cost of deploying and operating it, and the reliability customers could expect.

I think the useful way to follow this project is through those questions. Watch for sustained workload results, error behavior and working inter-satellite communication, rather than treating each successful prompt as evidence of inexpensive space compute. The first response makes the experiment more concrete; the next results need to show how much dependable work the system can actually deliver.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn complex launches and engineering updates into stories that developers and founders understand and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).