Google says it has run its Gemma 4 model on a data center-class Tensor Processing Unit (TPU) aboard a prototype satellite and received the answer back on Earth. The company’s 30 October update calls this the first response from its Project Suncatcher spacecraft. It is a limited test of onboard inference and communications, not an operational data center in space.
Google launched the prototype on 1 October 2026. During commissioning, its team powered up the host computer and TPU trays and ran early machine-learning tests, according to the update. In a ground-station pass this week, engineers sent a prompt to the onboard model asking for a short poem; the satellite returned a response while passing near Svalbard, Norway. Google says it subsequently tried other text prompts and prompts that returned payload telemetry. It has not published throughput, sustained uptime, power-use or model-quality results for this orbital workload.
Why the first response matters
A TPU is a processor designed for machine-learning workloads. Getting a model to execute on one in orbit and returning its output checks several pieces of a basic path: the payload can be powered, a request can reach it, inference can run, and an answer can come back. The demonstration does not show that many processors can run continuously or that satellites can share a large workload.
That distinction matters because Google’s earlier Suncatcher system design envisions solar-powered satellites connected through high-bandwidth optical links. The proposed architecture depends on close formation flying, reliable processors in the radiation environment and thermal control. Google’s 2025 research described ground-based radiation testing of a Trillium TPU and a laboratory optical-link demonstration. Those tests informed the design, but neither established how a production-scale constellation would behave in orbit.
The link requirement is especially demanding. In its 2025 research, Google reported 800 Gbps in each direction from a single laboratory optical-transceiver pair. The proposed constellation would need close-flying satellites and much greater aggregate bandwidth to split large machine-learning jobs across them. A bench result is useful design evidence, but it is not a measured inter-satellite link. The planned 2027 two-satellite test is the first stated orbital step toward checking that part of the architecture.
What remains unproven
Google says its next orbital tests will examine TPU behavior under thermal stress and radiation exposure. It plans to launch two interconnected satellites in 2027 to test another part of the concept. That timetable is a company plan, not a deployed service or a guarantee of performance.
For infrastructure operators, the useful signal today is narrower than the phrase ‘space data center’: one prototype has completed a first model request-and-response cycle. The harder questions are whether the hardware remains reliable over time, how efficiently it handles sustained work, and whether satellites can exchange data fast enough to run distributed workloads. Google has not yet supplied orbital evidence that answers those questions.
