Particle.news
Download on the App Store

Google Unveils Project Suncatcher to Test AI TPUs in Orbit by 2027

The effort pursues near‑continuous solar power in sun‑synchronous orbit to relieve terrestrial energy plus cooling constraints for rapidly growing AI workloads.

Overview

  • Google published a research plan for tightly clustered, solar‑powered satellites carrying TPUs and linked by free‑space optical connections in dawn–dusk low Earth orbit.
  • A learning mission with Planet targets two prototype satellites by early 2027 to test TPU operation in space and validate optical inter‑satellite links for distributed machine learning.
  • Bench tests demonstrated 800 Gbps each way on a single optical transceiver pair, while radiation trials on Trillium (v6e) TPUs showed irregularities only after ~2 krad(Si) with no hard failures up to 15 krad(Si).
  • Google’s models indicate tens‑of‑terabits‑per‑second networking could be reached by flying satellites hundreds of meters apart with modest station‑keeping, though thermal control, high‑bandwidth downlinks, and reliability remain open challenges.
  • Cost modeling projects launch prices below $200/kg in the mid‑2030s could make orbital compute cost‑comparable per kW‑year, with environmental benefits still debated given launch emissions and concerns over debris plus astronomy impacts; other efforts include NTU research, a Thales study, and Starcloud’s Nvidia‑powered test this month.