Mountain View, California, November 4, 2025- Google Research has publicly launched the framework for its bold new initiative, Project Suncatcher, proposing a future where solar-powered satellite constellations carry custom TPUs and high-speed optical links to build what may become the world’s first space-based AI infrastructure.
According to the blog post authored by Travis Beals, Senior Director of Paradigms of Intelligence, “Project Suncatcher is 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 system design envisions a network of small satellites operating in a dawn-to-dusk sun-synchronous low Earth orbit, where continuous sunlight can deliver up to eight times the energy availability of ground systems. Within that orbital band, the team says they can significantly reduce battery mass while enabling high-bandwidth compute operations.
One major engineering challenge is the “data-center-scale inter-satellite links”, the post describes. The team reports that “we have already begun validating this approach with a bench-scale demonstrator that successfully achieved 800 Gbps each-way transmission (1.6 Tbps total) using a single transceiver pair.”
The architecture would scale by flying satellites in extremely tight formations, with a cluster radius of approximately one kilometre, and adjacent satellites separated by an average of 100-200 metres. Models based on Hill-Clohessy-Wiltshire equations refined via JAX-based simulation suggest modest station-keeping is sufficient for orbit maintenance.
Radiation resilience is another focus. The team tested their Trillium v6e Cloud TPU in a 67 MeV proton beam and found that while high-bandwidth memory systems showed irregularities after about 2 krad(Si), no hard failures were observed up to 15 krad(Si), a margin well above the estimated ~750 rad(Si) dose for a five-year mission.
“Trillium TPUs are surprisingly radiation-hard for space applications,” the blog concludes.
From an economic standpoint, the study projects that if launch prices fall to below USD 200 per kg by the mid-2030s, the cost per kilowatt-year of orbiting compute could approach that of ground-based data centres. To begin validating the concept in practice, Google says it will partner with Planet Labs PBC to launch two prototype satellites by early 2027, enabling in-orbit testing of TPUs and optical links.
The blog notes that while the physics and early economics seem favourable, significant engineering hurdles remain, including thermal control, on-orbit reliability, and high-bandwidth ground links.
If Project Suncatcher succeeds, it could rewrite the definition of “the cloud” by placing a global ML-compute fabric beyond Earth’s surface.
.jpg)
