Google Launches First Orbital Data Center Satellite Running On‑Board TPU Inference

Google’s Suncatcher prototype reached orbit aboard a SpaceX Falcon 9 earlier this month and has completed its commissioning phase. The spacecraft carries a single Tensor Processing Unit capable of delivering up to 1 kW, but thermal limits restrict active compute windows to roughly twenty minutes per orbit.

Engineers sent the first inference prompt to the pre‑loaded Gemma 4 model, asking it to compose a haiku about the orbital view. The model responded with a concise verse, confirming that the TPU can execute large‑language‑model inference in the harsh space environment.

A key focus of the ongoing test campaign is radiation resilience. Ground‑based particle‑accelerator trials suggested minimal bit‑flip impact for inference, yet real‑world exposure may reveal subtle data‑corruption patterns that standard monitoring misses. The team will correlate telemetry with model output to quantify these effects.

If the current TPU proves reliable, Google plans to launch two follow‑up satellites in 2027 with higher power budgets and optical inter‑satellite links, enabling coherent distributed workloads across a constellation. Both this effort and SpaceX’s parallel vision depend on the rapid reusability of Starship to achieve the launch cadence required for economic viability.

Why it matters for GPU / AI infrastructure

Space‑based inference nodes introduce a new deployment tier for AI workloads, offering low‑latency connectivity to remote regions and reducing terrestrial data‑center footprint. Validating TPU performance under radiation and power constraints informs the design of future radiation‑hardened accelerators and influences capacity planning for hybrid ground‑orbit compute fabrics.