Google tests space-based TPUs with Gemma 4 model aboard Suncatcher satellite
Prototype satellite carries four TPUs, generates up to 1 kW, and runs inference with Google DeepMind’s Gemma 4, including poetry generation and orbital telemetry.
1 source · cross-referenced
- A prototype satellite named Suncatcher, carrying four Google TPUs and generating up to 1 kW of electricity, is now operational in orbit.
- The satellite runs Google DeepMind’s Gemma 4 model to perform inference tasks such as writing poetry and reporting orbital conditions.
- Engineers tested the satellite’s subsystems for three weeks before activating the model, which produced a haiku about being in space.
- Google plans to assess TPU performance under space radiation and thermal stress, with two follow-on satellites planned for 2027.
- The project aims to demonstrate feasibility of distributed inference workloads using inter-satellite optical links.
Google’s Suncatcher prototype satellite, launched this month on a SpaceX Falcon 9 rocket, is now online and performing inference using Google DeepMind’s Gemma 4 model. The spacecraft carries four Google Tensor Processing Units (TPUs) and generates up to 1 kW of electricity, though power constraints limit active TPU operation to roughly 20 minutes per session.
After a three-week commissioning phase to validate subsystems, engineers on the ground issued the first prompt to the onboard model, asking it to write a haiku about being in space. The model responded with: “Blue marble below / Data streams in silent grace / Starlight feeds my core.” Engineers also tested the model’s ability to report on orbital conditions around the satellite, with a company spokesperson confirming nominal operations.
The ongoing test campaign will evaluate how the TPUs perform under thermal stress and in the space radiation environment. Google has previously stress-tested TPUs in Earth-based particle accelerators, but space-based radiation can cause bit flips that corrupt data without triggering standard monitoring tools. Ground simulations suggested radiation impacts would be minimal for inference workloads, and the in-orbit tests aim to validate those predictions.
If the current prototype succeeds, Google plans to launch two additional satellites equipped with advanced TPUs in 2027. Those follow-on spacecraft are expected to support higher power generation and demonstrate inter-satellite optical links, enabling coherent distributed inference workloads. Google’s approach differs from some competitors by emphasizing high-speed links between closely flying satellites, whereas other companies plan to run inference on individual satellites.
The Suncatcher project is part of a broader push to explore space-based data centers, a concept that remains speculative and economically unproven. Alphabet, Google’s parent company, holds a significant investment in SpaceX, whose Starship rocket is expected to be required for deploying large-scale orbital data centers. A recent peer-reviewed analysis by Suncatcher researchers estimated SpaceX would need to fly Starship approximately 1,800 times to achieve cost reductions comparable to those seen with earlier vehicles.
- Nov 11, 2026 · Hugging Face
Hugging Face proposes token pooling-aware training to improve compressed multi-vector retrieval
Trust79 - Nov 11, 2026 · TechCrunch — AI
Fractile raises $720M Series C to develop AI chips and compiler stack
Trust79 - Nov 11, 2026 · The Verge — AI
Google’s Tabletop AI experiment lets users build custom collaborative apps via waitlist
Trust79