research.google via Reddit

Google Unveils Project Suncatcher, TPU Satellites by 2027

TL;DR

  • Google's Project Suncatcher plans two prototype TPU satellites with Planet by early 2027, in a 650 km dawn-dusk sun-synchronous orbit for an eventual AI compute constellation.
  • Bench tests of the free-space optical link hit 800 Gbps each-way (1.6 Tbps total); Trillium v6e Cloud TPUs showed no hard failures up to the max-tested 15 krad(Si) radiation dose.
  • The economic case depends on launch costs falling below $200/kg by the mid-2030s to make orbital compute cost-comparable to terrestrial data centers.

Google wants to put TPUs in space. In an exploratory design paper published on its research blog, the company unveiled Project Suncatcher, a moonshot plan for solar-powered satellite constellations carrying Trillium v6e Cloud TPUs in a dawn-dusk sun-synchronous low Earth orbit around 650 km up. Two prototype satellites, built with Earth-imaging operator Planet, are slated to launch by early 2027.

Travis Beals, Senior Director, Paradigms of Intelligence, frames Suncatcher as "a moonshot exploring a new frontier: equipping solar-powered satellite constellations with TPUs and free-space optical links to one day scale machine learning compute in space." The post places it within what Google calls its "long tradition of taking on moonshots that tackle tough scientific and engineering problems."

The illustrative design is an 81-satellite cluster roughly 1 km across, with individual spacecraft oscillating between about 100 and 200 meters of each other. Bench tests of the free-space optical link reached 800 Gbps each-way (1.6 Tbps total) using a single transceiver pair. On radiation, the team reports that Trillium's "High Bandwidth Memory subsystems were the most sensitive component," showing irregularities after 2 krad(Si), nearly three times the expected five-year mission dose of 750 rad(Si), with no hard failures up to the maximum tested 15 krad(Si). In orbit, Google argues, solar panels can be "up to 8 times more productive than on earth, and produce power nearly continuously."

The economic case rests on cheaper launches. Google projects that if costs fall below $200/kg by the mid-2030s, orbital compute could become cost-comparable to terrestrial facilities on a per-kilowatt/year basis. The paper's results so far are ground-based; whether the optical link holds between satellites drifting 100-200 meters apart in orbit is what the 2027 Planet flight is meant to answer.

The announcement lands in a dense stretch for AI infrastructure. Our tracker logged 443 stories over the last 90 days, mostly terrestrial buildouts; Suncatcher sits in a different category, a research mission on a decade-plus horizon.