Google Runs Gemma 4 on a TPU in Orbit for the First Time

Google Research has run an open-weight language model on a data-center-class tensor processing unit in low Earth orbit. According to the Project Suncatcher update , the team uploaded a prompt to its satellite, asked Gemma 4 to write a haiku about space, and downloaded the response during a pass over Svalbard, Norway.

The haiku exercised little compute, but it confirmed the complete inference path: a ground station sent the prompt, the orbital system activated the accelerator, the model generated text, and the satellite returned that text to Earth. Google described the October 30 run as its first AI response generated in orbit with a data-center-class TPU.

Four TPUs, one kilowatt

Planet built the prototype satellite, called MVP, and launched it on a SpaceX rideshare mission. It carries four Trillium-generation TPUs, an installed copy of Gemma 4, and solar panels that produce roughly one kilowatt of power. Google uses the same TPU family for production AI workloads in terrestrial data centers. Most spacecraft computers use lower-power processors engineered for radiation tolerance, making the orbital behavior of these accelerators a central research question.

From uplink to downlink

1. A ground station uploaded a prompt to the satellite.
2. The onboard system activated a TPU and passed the prompt to Gemma 4.
3. The model generated a response locally.
4. The satellite transmitted the output to the Svalbard ground station.
5. Later tests generated space-themed limericks and responses describing payload telemetry.

Google currently limits model runs to 15 minutes because heat management sets the duty cycle. Vacuum prevents convective cooling, so the satellite must conduct heat away from the processors and radiate it into space. Repeated transitions between sunlight and shadow also produce thermal cycling that terrestrial data centers rarely encounter.

One response validates the path

The October 30 test established that the satellite can receive a request, execute a transformer model on its onboard accelerator, and return coherent output. A single haiku provides no benchmark for throughput, latency, energy per token, sustained utilization, or long-term reliability. Those measurements will determine whether orbital TPUs can support workloads larger than brief inference jobs.

Radiation presents another unresolved problem. High-energy particles can alter values in memory or processor state, producing transient errors known as bit flips. Gemma 4’s accessible weights let researchers checkpoint the model, reload known states, and compare orbital outputs with ground-based reference results. Google plans to use later tests to measure error rates, recovery behavior, and hardware performance under continued radiation and thermal exposure.

Lasers determine whether it scales

Project Suncatcher’s proposed architecture extends far beyond the MVP satellite. Google has outlined a constellation at roughly 650 kilometers in a dawn-dusk sun-synchronous orbit, where each spacecraft would carry dozens of Trillium v6e Cloud TPUs. This orbit reduces time in Earth’s shadow and gives the solar arrays longer periods of generation.

The satellites would operate as a distributed cluster through free-space optical links, which transmit data with lasers instead of fiber. Google’s design targets bandwidth approaching 800 Gbps between spacecraft separated by roughly 100 to 200 meters. Maintaining that rate requires precise formation control and beam pointing because small alignment errors can move a narrow laser beam away from the receiver.

A separate pair of Planet-built prototype satellites is scheduled to launch by early 2027. Their primary task will be formation flying and inter-satellite link testing. Sustained bandwidth, pointing stability, fault recovery, and synchronization across moving nodes will determine whether the design can support distributed AI workloads.

No developer endpoint yet

Project Suncatcher has no public API, SDK, pricing, service-level target, or availability date, and developers cannot deploy workloads to MVP. The current experiment supplies flight data for accelerator reliability, thermal management, radiation tolerance, and orbital networking.

Several measurements remain necessary before the project can support a credible computing service:

• Sustained inference and training throughput
• Latency and energy consumed per token
• Thermal performance during long, highly utilized runs
• Radiation-induced error rates and recovery times
• Optical-link bandwidth under formation drift
• Launch, maintenance, replacement, and ground-link costs

The long-term proposal is a compute platform powered largely by sunlight and less dependent on terrestrial substations, cooling systems, and water supplies. Its practical workloads will also depend on data location: information produced by satellites could be processed nearby, while ground-based applications would incur the cost and latency of moving data into orbit and back. The haiku test gives Google one successful end-to-end run; the planned formation and endurance trials will show whether that result can become a durable computing system.