Google’s Suncatcher Satellite Runs Gemma 4 in Orbit, Opening Tests for Space-Based AI Data Centers

Google’s Project Suncatcher satellite has run an artificial intelligence model in orbit and transmitted its first response to Earth. The spacecraft used an onboard Tensor Processing Unit, or TPU, to execute Google DeepMind’s Gemma 4 model, which answered a prompt by composing a haiku about space: “Blue marble below / Data streams in silent grace / Starlight feeds my core.”

The demonstration followed a commissioning period of roughly 3 weeks after the satellite launched earlier in October aboard a SpaceX Falcon 9 rocket with more than 100 other spacecraft. Google also tested space-themed limericks and asked the satellite to report orbital conditions, according to coverage from The Verge and TechCrunch. A Google spokesperson told TechCrunch that the satellite was operating nominally.

Project Suncatcher is testing whether AI processors can perform useful workloads in the harsh conditions of orbit. The prototype carries a single TPU and can generate up to 1 kW of electricity, but its power budget limits the processor to roughly 20 minutes of operation at a time. That constraint complicates both computing schedules and communications with ground stations.

The immediate test involves more than writing poetry. Google is examining how the TPU and the satellite’s wider systems respond to thermal stress, radiation exposure, and the delays involved in space-to-ground data transfer. The Verge described the first response as evidence that the TPUs booted successfully, while TechCrunch reported that the satellite was also used to provide telemetry about its surroundings.

Radiation exposure is the central technical challenge in the next phase. Energetic particles can cause “bit flips,” changing stored values in memory and corrupting computations without activating ordinary monitoring systems. Google has tested its TPUs with particle accelerators on Earth, but orbital conditions provide the real operating environment needed to measure their effects.

Ground testing indicated that radiation would have a smaller effect on inference than on model training. Suncatcher’s current mission will test that distinction in practice by observing whether the TPU can run inference reliably over time. The results will help determine the amount of error detection, redundancy, and fault recovery that future orbital AI systems require.

Google’s next milestone is a pair of interconnected satellites planned for 2027, with TechCrunch describing plans to launch 2 additional spacecraft. Those satellites are expected to generate more power and demonstrate optical links between spacecraft, creating the foundation for distributed workloads in orbit.

That architecture separates Suncatcher from proposals that keep inference on individual satellites. Google’s design centers on closely flying spacecraft connected by high-speed laser links, allowing processors to share work across a network. SpaceX is also pursuing space-based data-center concepts and is expected to use laser communications, although TechCrunch reported that its satellites would orbit farther apart and focus less on tightly coordinated processing.

Suncatcher’s first transmission is a small demonstration with a large engineering question behind it: whether some AI processing can move closer to space-based data sources and operate without routing every workload through Earth-based infrastructure. The current satellite’s limited power and short operating windows show how far the concept remains from a commercial data center.

The project also places Google inside a growing competition over orbital computing. Alphabet has an investment relationship with SpaceX, while companies including Starcloud and Cowboy Space are pursuing different approaches to space-based AI infrastructure. Google’s emphasis on linked satellites makes networking, synchronization, radiation tolerance, and power generation as important as the TPU itself.

Google will continue testing the satellite’s TPU under thermal and radiation stress before deciding how the system performs beyond controlled demonstrations. The 2027 twin-satellite milestone will provide a more meaningful test of inter-satellite computing than the current single-spacecraft mission.

Success would give Google data on whether orbital processors can run repeatable inference workloads and exchange information through optical links. Failure would still identify limits in power management, radiation protection, or communications. For now, a generated haiku marks the first confirmed workload from Suncatcher, while the decisive results will come from sustained operation and coordination among multiple satellites.