Google's Project Suncatcher Puts TPUs in Orbit Aboard a Falcon 9
SpaceX's Transporter-18 rideshare launched Google's Project Suncatcher pathfinder on 1 October from Vandenberg, putting the company's TPUs in orbit for the first time. The satellite, built with Planet, will show whether AI accelerators survive launch stress, radiation, and vacuum heat. Google plans a two-satellite laser-link test in 2027 before any compute constellation.
On this page
130 payloads, one pathfinder
A Falcon 9 lifted off from Vandenberg Space Force Base on 1 October at 2:32 p.m. Eastern on SpaceX's Transporter-18 rideshare mission, carrying 130 payloads to low Earth orbit. The highest-profile one is a pathfinder spacecraft for Project Suncatcher, Google's plan to move AI compute into space, built in partnership with satellite operator Planet, and it is carrying the company's TPUs on their first flight above the atmosphere. Sundar Pichai framed the experiment plainly: can Google's TPUs survive and operate in space? The company is going to find out.
The premise is energy arithmetic rather than science fiction. Low Earth orbit offers near-constant sunlight, up to eight times the solar power available to a ground data center, and Google's research group argues that compute scaled on Earth will eventually run into land, power, and cooling limits that orbit does not have. The pathfinder does not test any of that economics. It tests the part that comes first: whether the hardware survives the ride and the environment.
What kills a chip in orbit
Three failure modes dominate, and the pathfinder is instrumented for all of them. Launch loads expose a spacecraft to around 10 g with individual components seeing far more, and Google's ground tests at UC Davis's proton beam facility found that Trillium TPUs running AI workloads held up remarkably well, absorbing a total ionizing dose beyond what a five-year mission would deliver without failing. Radiation is the quieter threat: high-energy particles flip bits and degrade silicon over time, which is why the team monitored for errors while the chips were actively working.
Heat is the third problem, and it has no convection to help. With no airflow in vacuum, a working TPU sheds warmth only through carefully designed paths, so Google validated a combination of heat pipes and radiators in a thermal vacuum chamber before flight. The orbital results from all three tests feed directly into the next design, and the company is explicit that this flight is about finding points of failure rather than proving the concept.
The 2027 laser problem
Even if the silicon survives, the architecture's hardest constraint is optical. Google's plan is clusters of satellites, each carrying dozens of TPUs, networked by lasers; a 2027 mission with two satellites is meant to demonstrate inter-satellite links at the bandwidth machine-learning workloads demand. Today's space lasers are optimized for low bandwidth over long distances, while Suncatcher needs very high bandwidth between spacecraft that are close together and both moving, with pointing precision Google compares to hitting a coin-sized target from miles away.
For the local AI audience, the honest read is that this is a decade-scale bet, not a product. Nothing launched this week brings compute closer to your device; it explores whether the supply side of frontier models can keep growing when grids and permits stop cooperating. The interesting near-term output is engineering data: if TPUs shrug off a five-year radiation dose in ground tests and the pathfinder confirms it in orbit, the cost curves that keep frontier compute scarce look different than they did last month.